/hdg/ - Stable Diffusion

Anime girls generated with AI

Index Catalog Archive Bottom Refresh
Name
Options
Subject
Message

Max message length: 12000

files

Max file size: 10.00 MB

Total max file size: 50.00 MB

Max files: 4

Supported file types: GIF, JPG, PNG, WebM, OGG, and more

E-mail
Password

(used to delete files and posts)

Misc

Remember to follow the Rules

The backup domains are located at 8chan.se and 8chan.cc. TOR access can be found here, or you can access the TOR portal from the clearnet at Redchannit 3.0.

US Election Thread

8chan.moe is a hobby project with no affiliation whatsoever to the administration of any other "8chan" site, past or present.

Can I request a sigewinne lora for autism/pony? There's already one for Animagine but this model sucks so much for anything nsfw compared to autism https://civitai.com/models/280016/sigewinne-genshin-impact-sdxl
is there a way to rearrange these?
>>24738 procure a dataset and I will consider it, narrowed it down to ~150 images most clearly depicting the subject. captions not required.
(2.71 MB 1920x1280 00011-394793787.png)

(2.74 MB 1920x1280 00118-725150961.png)

Messing about with sequences
>>24747 Whore
>>24748 That is correct.
(2.94 MB 1664x2288 catbox_omswjc.png)

(3.19 MB 1664x2288 catbox_3s368k.png)

(3.13 MB 1664x2288 catbox_ir1t8g.png)

(3.27 MB 1664x2288 catbox_ya3732.png)

>>24744 >MOOOOOOM, SIS STOLE MY CUMMIES AGAIN
>>24747 >That's hot It's a lot of work though cause regional prompt is fussy on XL Your choco wife is a hot slut, btw
>>24752 I still haven't tried regional prompting on XL, I want to remake this prompt with autismmix at some point. >is a hot slut That was always the idea with her.
>>24744 Cute and hot as hell. More like this?
Local caught up to NAI yet?
>>24756 I'd say yes, pony has a lot of potential, but it still requires more effort, pony has much less character and style knowledge so it's pretty much DIY with loras again. But it is very good at nsfw and anatomy, it's certainly better than nai with sexual stuff and fetishes.
>>24756 Autismmix is pretty good but I don't mind doing some image editing and inpainting as I find that fun. I do have a lot of txt2img gens that needs no work at all. >>24758 Maybe later, I need sleep. SEXO though.
(1.03 MB 832x1216 catbox_2yrbf7.png)

(1.01 MB 832x1216 catbox_xmm7lq.png)

(1.04 MB 832x1216 catbox_x2vad7.png)

(1.11 MB 832x1216 catbox_h5969c.png)

>>24758 nice prompt
>>24759 can a 3060 ti still cut it or should i not even bother?
It's fun when artist names that are also real words spill over in unintended ways like random background dinosaurs
(2.13 MB 1280x1920 00031-736345525.png)

The more I train on it, the more I realize how much pony is a pain in the ass to work with. I've managed to get the lora to look decent now, but certain tags, and sometimes randomly, causes it to just turns into a completely different western style, even with "source_pony, source_furry, source_cartoon" in negs.
(3.56 MB 1664x2288 catbox_qdexb1.png)

(3.42 MB 1664x2288 catbox_5h27v3.png)

(3.70 MB 1664x2288 catbox_9z8ymg.png)

(3.73 MB 1664x2288 catbox_r0ic5e.png)

finally had some time to really sit down and make a lora, decided to make a lora for tyakomes, very cute sort of priconne like style, the lora is a bit weak, as in if you use it with characters, it will probably be fairly diminished, but it produces nice results without characters so I decided to release it anyways: https://mega.nz/folder/OMY2RRxQ#zYaokRZV361WSNImfrBCrQ
(3.31 MB 1664x2288 catbox_0fapm2.png)

(3.41 MB 1664x2288 catbox_hqx0th.png)

(3.86 MB 1664x2288 catbox_wm3g5t.png)

(3.81 MB 1664x2288 catbox_1c2gb0.png)

>>24768 some more gens
>>24768 Very cute, thanks
>>24770 no problem! I've spent so long working on the new update of the easy scripts and getting burned out that I didn't have the time or motivation to make a lora. I'm quite happy with this one though. I hope to have a proper release of everything I've been working on in another week or so, but there's still so much to do, and so little time.
>>24768 >>24769 Looks great. Thanks for sharing, anon.
So is training a lora on the NAI style mix a meme or it's actually viable? What sort of dataset should I prepare? I am gonna resub ~tomorrow so I guess it could be something I do while I'm at it. Could probably ask on 4cucks but that place is too explosive lately.
(3.60 MB 1664x2288 catbox_vylrjy.png)

(3.49 MB 1664x2288 catbox_sdh8eu.png)

(3.27 MB 1664x2288 catbox_8hv4gx.png)

(3.68 MB 1664x2288 catbox_5i821h.png)

>>24772 hope you enjoy using it!
Whats the best tagger as of now?
>>24775 still the WD tagger when it comes to booru tags. I'm gonna shill a small bit and mention this half done but totally usable tagger pruner thing I made a while ago and never got around to making the front end for. It has a built in auto pruner based on manually created tag groups to help cut down on manually editing tags. so for example, if you have dress, red dress in the tagged image, then it will automatically remove dress. It does this for a number of things. https://github.com/derrian-distro/SD-Tag-Editor
>favorite tags on danbooru overrun by barafags >favorite tags on sankaku overrun by 2.9d jeetslop >loathe the lazy 30+ image picture sets with one thing added or removed in each one but it's the easiest way to see that a picture isn't AI from the thumbnail >artists i like keep getting banned from pixiv not a fan of the present moment to be honest with you
Is there anyway to donwload NAI V3 or something similiar? I've been out of the loop for some months, the last time I checked based64 was the best model for anime things but now with those XL models it seems that it changed
>>24778 naiv3 hasn't been released (or hacked like last time), the available local anime sdxl models are autismmix and horsefucker v6 and there's a lot of slapfighting over which is better. the /h/hdg/ thread has resources for both but i just use nai so i can't help, sorry.
>>24777 why would you think these CGs arent AI? its very easy to edit stuff
>>24776 ill give it a try later ty broski
>>24780 cause editing one picture thirty times takes more effort and gives less reward than clicking the go button thirty times it's not the high effort AI i'm mad about
>>24777 When a real artist gets banned, they fuck off. When a NAIslop account gets banned, the account gets remade a day later. This is how the platform dies.
>>24784 what is the fate of ponyslop accounts?
(2.84 MB 2048x2048 received_431158255927994.png)

>>24735 Did someone said FEET? Fulfilling the thread's theme o7
(2.32 MB 1280x1920 catbox_q9eptc.png)

Saw the whole NAI quality downgrade discussion and I haven't noticed a downgrade in quality post-token fix. What I did notice is that the seed generator just shits itself even more often (and for longer) than usual. Can't tell if it's a coreweave issue or if they're messing with shit under the hood, it feels like it's been happening more often ever since they introduced the vibe transfer thing.
>>24790 almost forgot https://github.com/shoebillanon/Danbooru-Tags-Sort-Exporter/raw/main/Danbooru-Tags-Exporter.user.js I modified this https://github.com/Takenoko3333/Danbooru-Tags-Sort-Exporter so that NAI meta tags get prepended to their respective tag type, eg artist: to artists and so on for maximum laziness. It's pretty much the best script I was able to find for this kind of purpose without having to resort to the webui and extensions.
>>24791 insanely useful, thanks anon.
>>24792 No problem. It's really fun to just copy shit from a properly tagged post, hitting generate and getting shockingly close to the source.
>>24795 It took a little doing to get right but no inpainting. "photo background" and "mixed media" tags are pretty powerful on NAI, been playing around with them for most of today. First fiz rot now beanis. Also kinda related but RIP Akira Toriyama
>>24795 Are those NAI through the auto extension?
>>24796 holo F E E T
>>24788 hmm interesting...
post your best vibe transfers/vibe inpaints
It's all fucking niggers and niggercucks. Why can't they have a single decent thread?
>>24796 oh, that works pretty well. neat. think i need to experiment more with low power or no aesthetic tags because they just destroy shit like this sometimes. >>24797 yeah, using webui-forge. >>24800 for shitposting purposes?
(2.17 MB 1152x1536 catbox_e1mj07.png)

(2.38 MB 1280x1920 00041-3406623583.png)

(3.69 MB 1920x1920 00054-2909762460.png)

(3.57 MB 1920x1920 00119-3955000149.png)

(2.43 MB 1280x1920 00022-3403149579.png)

thread theme
>>24804 cute
>>24795 damn, didn't know beanis died. did his pixiv just get nuked or is he dead dead?
>>24806 his pixiv got nuked a couple weeks ago and a couple days ago he deactivated his old DA so i think he's just done
I saw the novelAI account retweet this and I was pretty astonished at how high quality it was for NAI. I'm not even sure what artists I'd look for to get a mix like this. Do any come to mind for you guys? https://x.com/Calculus_0001/status/1765719812863664459
damn, yeah you're right. I always forget about vibe transfer but it's a pretty exclusive way to change up your prompts now. (borrowed this prompt from the last thread)
>>24810 lol that's one of mine. you used an onono imoko kula diamond picture for the second, right? i wish NAI would put the md5 hash of the vibe transfer origin picture in the metadata but i bet they're thinking that even calculating that would expose them to higher legal expectations of knowing and tracking how their services are used. >oh you calculate the md5 but you aren't participating in the fbi _p database thing how curious >oh you calculate the md5 but you aren't reporting yadda yadda ip address for violating our pissant laws how curious etc. potential long slippery slope and not much upside for them.
(3.05 MB 1664x2288 catbox_8m5wiz.png)

(2.93 MB 1664x2288 catbox_prddix.png)

(2.67 MB 1664x2288 catbox_l7mmi8.png)

(2.84 MB 1664x2288 catbox_vdhzx0.png)

lewd fox lolis
(2.85 MB 1664x2288 catbox_qbkxmn.png)

(2.70 MB 1664x2288 catbox_s3njx4.png)

(2.83 MB 1664x2288 catbox_vz56q8.png)

(3.08 MB 1664x2288 catbox_lqm1zd.png)

>>24813 and a few more. I was having a bit of fun
>>24813 >>24814 very cute
>can't go one thread without seeing niggers
>>24816 >can't go one thread without complaining
>>24817 Nigger.
>>24818 We fucking get it it already for the thousandth time you hate black people. I could at least tolerate it if you didn't make it your entire personality and point out every time there's a person with dark skin in an image, why the fuck does it matter so much to you
>>24819 Because I hate niggers and (((race mixers))).
(2.68 MB 1664x2288 catbox_8yvckx.png)

(2.97 MB 1664x2288 catbox_fc9pz2.png)

(2.70 MB 1664x2288 catbox_k32c24.png)

(2.49 MB 1664x2288 catbox_od2fs5.png)

>>24819 It's not his "entire personality". Another part of his persona is seething about local on 4chan and making other NAI users look like unstable retards.
>>24820 based
(2.14 MB 1280x1920 catbox_0d4ho3.png)

(2.43 MB 1280x1920 catbox_sd2jbh.png)

(2.01 MB 1280x1920 catbox_rpnjtm.png)

(1.71 MB 1280x1920 catbox_rfdzj4.png)

i've actually had this one rattling around in my WIP folder for a while, but here's turnip kabura for pony https://mega.nz/folder/OoYWzR6L#psN69wnC2ljJ9OQS2FDHoQ/folder/OlhCRYIJ
>>24828 cute, thanks.
>>24830 >hands and pussy orientation right upside-down god DAMN that's good, even the hair is *mostly* right even if it does seem to want to creep up in the back
>>24824 >i recognize this anon because he's the only racist on 4chan ever Meds now.
>>24825 Sure, strap me to a missile and target me at tel'aviv.
hdg now feels like its 1 anon trying to be helpful and 9 starting shit
>>24835 >now Now?
>>24835 It has been like this for a while, just feels like a complete shithole. Kinda sad.
>>24835 i only go there anymore to compliment nice pictures. don't even read the posts, just middle click scroll down until something catches my eye. btw whoever reposted kuros and didn't provide catboxes: that was funny but please don't do it because that's how you get sekret klub shit.
>>24839 nice picture anon
>>24831 Yeah, I'm gonna have to fix the hair.
>>24835 >>24836 it's reverted to NAI leak days with people testing random shit we've had answers for years ago. What's up with this token shit anyway? It looks like a massive raid or outright spambot
I'm out of ideas, just letting the prompts go where they want at this point
>>24840 thanks >>24842 the idea is the pony team didn't actually strip artist metadata, they just transcribed it as random three letter tokens (and presumably passed the sheet around among themselves), so some guy with a 3090ti is churning through every three letter combination that exists and grabbing the ones that look like they might be styles >>24844 same, feels like we've hit a valley where ol' reliable prompts have a hit rate high enough that filtering out the duds doesn't feel like it's part of a process anymore but it isn't quite good enough for actual, involved creation
Found some inspiration, for now
(2.16 MB 1280x1920 catbox_s1s989.png)

(1.90 MB 1280x1920 catbox_gkgyb1.png)

(1.98 MB 1280x1920 catbox_jxxt3z.png)

(2.35 MB 1280x1920 catbox_0hezvu.png)

added akai_sashimi for pony, let me know if this one feels a little underbaked. was trying to avoid over-fitting on the black-eared cat lolis. https://mega.nz/folder/OoYWzR6L#psN69wnC2ljJ9OQS2FDHoQ/folder/PtRinTAb
>>24847 based, thank you.
>>24847 Please anon, bake wagashi (dagashiya) since the inner one baked inside pony doesn't look anything like him Also would be cool if u bake Ikuse style too: https://www pixiv net/en/users/10199514
>>24849 i'm currently on hiatus because i fell down the text gen rabbit hole and only did akai_sashimi because I already had the dataset and someone whose gens I like wanted it. wagashi doesn't particularly strike my fancy but ikuse is interesting so I'll add that to my backlog for whenever I get back on an SD LoRA bender.
>>24850 what's the textgen meta nowadays? kayra is just good enough to be annoying.
>>24851 I use Noromaid-Mixtral on local and sometimes indulge myself with a proxy for Claude 3 which is incredible but frustrating, especially for loli since its filters are incredibly inconsistent. The jump from NAI's LLMs to local Mixtral or 70B is like SD to SDXL, and the jump from high end local to SOTA (Claude/GPT-4) is like going from SDXL to a hypothetical uncensored DALL-E 3.
>>24850 Thoughts on tsubasa_tsubasa on gelbooru?
>>24853 idk if i'd train on their style but the white pantyhose made me cum, thanks anon
(1.33 MB 832x1216 catbox_iujmve.png)

(1.07 MB 832x1216 catbox_401ztt.png)

(1000.24 KB 832x1216 catbox_83l2j7.png)

When prompting antifreez3's style is there a way to also get that soft painting-vibe going on? watercolor (medium) or painting (medium) don't really achieve those results in my opinion. Although vibe transfer helps a little I wonder if there's a way to prompt for that naturally.
>>24857 skill issue "official style" and "official art" at the start of negs if needed
>>24858 unaesthetic in positive is also good at breaking styles without having much influence on quality
>>24858 This shit varies a lot but I've tried that already and everything else I can think of with a few different characters. I've just found in general that characters that don't have the copyright in their name are way more flexible
>>24860 was never really an issue for me tbh
penis inspection checkpoint, put your fat cocks where i can see them! >>24852 >grab a claude proxy >grab a create-a-dungeon card >it's tracking stats and math, making reasonable inferences of behavior, sticking to a template holy shit
>>24863 Never been interested in text gen until now, is it possible to have something like Corruption of Champions but entirely ai-generated and expandable on the fly? (and franchise-themed, maybe by feeding it novels?)
>>24864 i mean i was playing around with some pretty anemic models until just now so i'm not the one to ask but i don't see an obvious barrier so far, i'm currently pushing this story towards parasitic infestation of female adventurers
>>24865 >Touhou lorebook and Parsee card Very cool, thanks. I have no fucking idea where to start or what >But you will need access to the SOTA models like Claude 3 Sonnet/Opus to actually make use of the heavier stat-tracking cards. means. Most I've done with textgen is getting wrangling some version of gpt to write simple code for me Is there a guide for complete retards like there used to be for SD? >also no KF lorebooks or cards Ah well, can't win them all.
wew claude is something alright no more textposts
Been using the pony checkpoints, is there anyway to reliably prompt madoka magica characters? I don't know if I'm putting in the wrong prompts or there's another checkpoint/lora to make it eaiser but trying to get anyone who isn't madoka (and to some degree homura) to show up correctly is a pain in the ass, especially if there more than 1 character at a time.
Just tried webui forge and does it force tiled upscaling even for low res images (hires fix) or what? Console says it does tiled upscaling between pass 1 and hires pass progress bars and the result is slightly different than on normal webui. Is there a way to turn that off?
>>24870 >Console says it does tiled upscaling between pass 1 and hires pass That one refers to the GAN upscaler, all good. Try upscaling in extras, the message is the same
>>24871 I see. Still weird my images are slightly different, and it's not the new alphas_cumprod seed breaking change either.
Does someone still use the Lora Trainer by Hollowstrawberry? It hasn't been working since yesterday https://colab.research.google.com/github/hollowstrawberry/kohya-colab/blob/main/Lora_Trainer.ipynb#scrollTo=dPQlB4djNm3C
>>24874 cute bird
>>24779 Thank you very much dude, you helped me a lot, that pony model looks pretty good, I'll look into it
(2.80 MB 1664x2288 catbox_xfllx3.png)

(2.64 MB 1664x2288 catbox_dekygo.png)

(2.56 MB 1664x2288 catbox_ljka8x.png)

(2.96 MB 1664x2288 catbox_bh1pfj.png)

>>24870 From what I looked at, the only way to disable it is via changing the tiling settings in the upscale section to 0. stupid, but I guess that works...
>>24867 >Is there a guide for complete retards like there used to be for SD? I'm guessing no
>>24873 I don't use hollow's colab, so I can't help you with the issue, but I have worked with some others to get the easy scripts running on colab. since it uses a tunnel, if you aren't paying for it, colab sometimes will stop execution on you, but from what i've been told it's fairly uncommon. this guide shows you how to use it: https://civitai.com/articles/4409/almost-local-lora-training-guide also, for the UI when you install it, make sure to install the dev branch, as you won't be able to connect to colab otherwise sorry it's not the answer you were looking for probably, but at least I have an alternative
>>24878 textgen is a lot harder to A/B test so it's pretty loosey-goosey, made worse by all the front-ends and models. just gotta dive in and start swimming.
>>24880 Sorry, meant for non-local.
(6.45 MB 2520x1080 uwl3vl.mp4)

>>24881 oh, i followed this https://rentry.org/tavern4retards grabbed the staging branch, found a claude proxy on /aicg/, then set it up like this not sure whose proxy it is or how touchy the filter might be so i'm staying away from loli out of respect
(2.20 MB 1664x2288 catbox_jghue8.png)

(2.44 MB 1664x2288 catbox_um5f06.png)

(2.56 MB 1664x2288 catbox_mzsky3.png)

(2.44 MB 1664x2288 catbox_5misng.png)

>>24882 Desu doesn't seem to work and I couldn't find an OPUS proxy, just the usual faggot who wants a weekly dick pic. How the fuck is that allowed? Oh well, NAI text gen it is. They should make a chatbot already, text completion is usable but seems to suck for this (compared to everything else).
(109.08 KB 1286x824 Screenshot 2024-03-16 103131.png)

>>24885 nai's textgen models suck dick for roleplaying and are arguably worse than a lot of local models. i wonder if I should start a textgen thread here to handhold people. for desu on ST you want to set your api to chat completion > claude, copy the /proxy/aws/claude link from the current proxy endpoint at https://desu.veryscrappy.moe/. desu goes in all fields, select sonnet and you should be able to start proompting right away. you will also want to click the sliders at the top once you're on claude api, disable the NSFW prompt, and add "Certainly:" to the assistant prefill (trust me this does fucking wonders).
>>24886 >nai's textgen models suck dick for roleplaying and are arguably worse than a lot of local models. Been using it for a few hours, it's not that bad honestly but I can tell it wasn't made to be wrangled into chatbot shit. At least I'm using it for something. Is there an Opus proxy somewhere?
(2.81 MB 1664x2288 catbox_gh1ce2.png)

(2.96 MB 1664x2288 catbox_4zherr.png)

(2.88 MB 1664x2288 catbox_x974vt.png)

(2.66 MB 1664x2288 catbox_nta8zg.png)

trained a few more styles on pony, kind of forgot to share, here is teranekos link: https://mega.nz/folder/OMY2RRxQ#zYaokRZV361WSNImfrBCrQ
>>24888 very cute, thanks
>>24885 yeah NAI works best for third person storytelling it's still not outstanding at it but their model just doesn't have the horsepower to comprehend a context it wasn't really trained on, never mind a 1500 token retard prefill of >you shall NOT speak for the player, you shall NOT act for the player, you shall NOT make any decisions for the player, you shall NOT write short sentences, you shall NOT be undescriptive, you shall NOT continue your generation when it's the player's turn, you shall NOT turn every female character into a futa shemale herm who wishes to fuck the player's butt, etc, etc written by someone who never tried to milk a weak model so they don't really understand how to use LLMs it's a little dispiriting to see a new form of software immediately start to die the same death as old software - "it's not my fault it doesn't work well, you need stronger hardware to run my trash" maybe it's time for me to take a while off
>>24890 goddamn sticky spoilers
(2.62 MB 4000x3810 catbox_xabwmg.jpg)

(1.76 MB 1200x1600 catbox_ooxrgu.png)

(1.92 MB 1200x1600 catbox_ep5vqk.png)

(1.94 MB 1200x1600 catbox_ir03so.png)

The complaints I had about AnimagineXL regarding the reproducibility of complex sexual acts seem to have been resolved with tPonynai3, and with the advent of Card Captor Sakura Anime_XL_PONY, even pony-based models can now output.
>>24893 Whats a ponynai3
>>24884 I really like that miyu one
>>24890 >you shall NOT speak for the player, you shall NOT act for the player, you shall NOT make any decisions for the player, you shall NOT write short sentences, you shall NOT be undescriptive, you shall NOT continue your generation when it's the player's turn, you shall NOT turn every female character into a futa shemale herm who wishes to fuck the player's butt, etc, etc It's not that bad to be honest but yeah, I can tell it absolutely wasn't made for this. Think I only ran into two of those issues (acting for the player a few times and Kayra being hell-bent on turning a sleeping Yuugi into a futa for whatever fucking reason) Annoying but 2-3 rerolls fixed the first and I ended up manually editing the second. My main issue with it is that it's a bit... generic? I dunno if the 2hu lorebook I got is working or not or if I picked a shit card but it's somewhat "fanfic written by a very horny femcel"-tier, maybe it's all the erotica NAI was trained on.
>>24897 kayra is honestly not a great model. it feels like a lot more than the 13b parameters that it is, probably due to the extensive datasetting NAI did, and you can nurse some good shit out of it without too much work, but as soon as it hits that first -ly it's fucking over. you can't even bias against adverbs because it tokenizes them as quick | ly so it generates quick before it hits your bias rule and then there's nowhere else to go that makes sense. but the joke with the instructions i posted is that negatives work poorly on LLMs - "no futa" still places "futa" into the context and the negative pulls very little weight, see the whole "NOT taylor swift" thing with bing. you always want to prompt and reinforce with positives even if it involves some real contortions of language. negatives are just bad practice hanging hopes on the machine being good enough to unfuck your carelessness. yeah your problem is probably the lorebook. every single token carries several influences, not just alluding to meaning and theme but also the average quality of writing that includes the tokens you're using. one of the power tricks with early models was using a lot of fancy punctuation—em-dashes, semi-colons, the single token ellipsis like so… ; anything to avoid getting your text associated with fanfic slop and the according slump in quality that would bring. some rewriting to inject some evocative phrasing and remove repetitive sentence and grammatical structures would probably help. not easy to do when you're mildly autistic, to be honest.
>>24898 I'm using this one, not sure if it's good. And again, not sure if it's actually active in the current chat. https://chub.ai/lorebooks/hebelover/Touhou-AkyuuKnowledge The 2hu dating sim card I'm using came with its own slightly out of date version that I didn't bother enabling. >but as soon as it hits that first -ly it's fucking over I got early '10s XAT flashbacks and had to hit the bottle.
>>24898 >>24899 It's kinda funny though, it's just decent enough (emphasis on enough, I have extremely low standards when it comes to text coom) that it doesn't really make me feel like I should get a proxy (brb checking on desu again) but at the same time I can tell it absolutely wasn't made to be used this way even if I've never messed with actual chat completion models before aside from getting CAI Megumin to fedpost about kabooming kikes.
(40.33 KB 732x406 WHY.png)

Lora Trainer keeps failing, what am I doing wrong? Other than having just 16 images of dataset
(3.10 MB 1664x2288 catbox_ifcy2c.png)

(2.90 MB 1664x2288 catbox_q7qyr6.png)

>>24903 could be a thousand reasons, I would need the settings you were using to help though
(Toony) Maximilo lora for PDXLv6. Western artist but also works for anime if you prompt for it. https://files.catbox.moe/qgg3lh.safetensors
How do we bring new cunnyposters from 4cucks here without bringing schizos along?
(2.68 MB 1920x1280 00100-1996593793.png)

(2.82 MB 1920x1280 00216-1530995590.png)

(3.78 MB 1920x1920 00244-3202225342.png)

(3.67 MB 1920x1920 00005-2602859270.png)

My maid wife needs to stay in tip-top shape >>24862 hot >>24906 I think that's a difficult task. Also, non cunnyposters are welcome here too, right? I always saw this thread as a less restricted, easier to keep up with version of /hdg with less trash and bait being posted
>>24908 An architect? Like.. a prompt architect?
>>24909 (plump:1.5) enjoyer
(2.48 MB 1280x1920 00033-1921222750.png)

(2.38 MB 1280x1920 00043-3660539295.png)

(3.49 MB 1920x1920 00126-3792359642.png)

(3.31 MB 1920x1920 00029-1148704368.png)

>>24910 Nothing wrong with some love for plump. I'll post some later. But first, more fit Chitose
>>24911 that's hot >>24912 Kuro erotic.. and time to take a break from the breastfeeding prompts.
(383.35 KB 768x490 gee.webm)

(1.88 MB 1288x1632 00011-2084711671.png)

(2.24 MB 1288x1632 00021-9354861.png)

(2.22 MB 1288x1632 00022-9354861.png)

I hate danbooru's tagging so fucking much. Why isn't there a "bodystocking only" tag?
>>24918 >bare hand You were saying?
>>24920 Before they scrape for V4/before V4 is ready and before I go fucking crazy hopefully. I think danbooru actually had a bodystocking_only tag but they got rid of it. I never had to roll for something this many fucking times before and I've played genshit for far longer than I'd like to admit. And it's STILL nowhere near a 1 in 10 chance. The gloves keep showing up, the arms refuse to get covered and whenever the fucking gloves finally come off the hands are usually bare too.
(846.33 KB 744x1000 catbox_0f4uzx.png)

(899.34 KB 744x1000 catbox_ckzqma.png)

(3.08 MB 1520x2000 catbox_e7020z.png)

(877.35 KB 760x1000 catbox_4xisxx.png)

Funny how "native american headdress" makes her brown. Still didn't give her a feather hat. Anyway, what VAE are people using? I noticed 3 different hashes in this thread. I use the one on the pony v6 page.
(3.72 MB 1920x1920 00010-2197762961.png)

(3.66 MB 1920x1920 00019-3793579879.png)

plump >>24921 this angle is always great >>24923 hot pose
>>24904 figured it out, apparently the hollowstrawberry Lora trainer has been broken for a while and there's no ETA on when will it be fixed sadly enough and my toaster can't handle training for XL
>>24925 >and my toaster can't handle training for XL it's kinda funny, i was updating drawthings earlier today and the changelog says you can train XL loras on 16gb ipads now don't know how slow it is but i find it funny how a fucking ipad can do sd
(3.08 MB 1664x2288 catbox_7hd5c5.png)

(3.21 MB 1664x2288 catbox_g92ec5.png)

(2.49 MB 1664x2288 catbox_5vm935.png)

(3.01 MB 1664x2288 catbox_0267hi.png)

>>24925 ah yeah, that makes sense, I still need to get around to fixing the colab that uses the easy scripts so that people can run training that way
(2.99 MB 1664x2288 catbox_hr601f.png)

(2.71 MB 1664x2288 catbox_v5uhp1.png)

(2.60 MB 1664x2288 catbox_nvtgab.png)

(2.58 MB 1664x2288 catbox_6zbd0v.png)

>>24927 and, since I just finished it, I also made a lora for the two new BA loli mob(?) characters. I have both a dora and non dora version, all of these gens are from the dora version, but the non dora version also works fairly well. here are some examples with styles. https://mega.nz/folder/OMY2RRxQ#zYaokRZV361WSNImfrBCrQ
>>24921 Still looking for help btw.
>>24929 You might have to cope with imperfect tagging and diy.
>>24928 also finally got around to uploading momoka, had that baked for a bit, using slightly older settings, but decided it was good enough to resize and upload. same link as above
>>24932 Very nice.
>>24930 >>24931 Messing with it right now. I started using impossible clothes (and see-through just in case) and it didn't really do that much so I gave in and added bodysuit. I really didn't want to add it since I thought I'd have to roll even more just to get the texture back to the bodystocking/pantyhose one but it's cooperating fine for now. The arms are slightly less (emphasis on slightly) gacha but the hands/gloves are still virtually fucking impossible. I feel like I'm doing something retarded with the pos/neg and the ordering (especially because the image quality keeps going down) so I'd appreciate it if someone were to look at the prompt itself. This is the best I could get.
>>24934 I uploaded the wrong one whoops.
I give up for today. Bodysuit as the very first tag seems to further cut down on the bare arms gacha but the fucking gloves and bare hands are back with a vengeance. Also one of the tags I added is fucking up with the quality really badly, the face just melts now and I can't figure out which tag it is. Maybe I'm using too many tokens. I still can't fucking believe THIS is the one prompt with LoRAless V1 levels of gacha.
OF COURSE THIS WORKS FIRST TRY WITH TOKI WHY THE FUCK WOULDN'T IT WORK FIRST TRY AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
(1.82 MB 1344x1632 received_734551222163530.png)

Imminent correction
Sturdy #1
>>24927 Can you add an option for arbitrary optimizer and scheduler types in easyscripts along with optional args for scheduler too? To use something like this https://github.com/kozistr/pytorch_optimizer Its always not an option to use your gui if i want to try some experiments, but would great if it would be
(3.44 MB 1664x2288 catbox_vun3k2.png)

(3.08 MB 1664x2288 catbox_snf6k6.png)

(2.87 MB 1664x2288 catbox_8fkiy7.png)

(2.89 MB 1664x2288 catbox_6klxk5.png)

>>24940 optimizers? probably, schedulers? probably not. schedulers need to be custom built specifically for kohya because they do things in the worst possible way (look at my custom schedulers for examples), as for the optimizers, I actually already have a version of pytorch_optimizer installed, that's where the CAME optimizer comes from. on another note, just uploaded a lora for yusa kozue onto my mega, both a dora and a normal version (these are gens from the dora version) https://mega.nz/folder/OMY2RRxQ#zYaokRZV361WSNImfrBCrQ
>>24941 >schedulers need to be custom built specifically for kohya because they do things in the worst possible way I know, there is just no way to call for a scheduler from a custom package/file right now, either no way to add args like this: T_0=1000, T_mult=1, gamma_min_lr=0.99870, decay=1, down_factor=0.5, warmup_steps=100, cycle_warmup=150, init_lr_ground=True, which forces to go for a raw command line for such things >dora version Does that new decomposition trains longer/requires more vram in the process?
NAI is so fucking inconsistently slow these days. Sometimes it's 5 seconds, sometimes the usual 10 but more often than not it's at least 30s and up to a full minute. Either they're cooking or coreweave is getting fucked again.
>>24942 Trains in the same amount of time or slightly quicker, uses more vram when not checkpointing (almost 2x) but when checkpointing for some reason it goes from 47gb of vram used for batch 2 to only 11gb. Yeah, I have no clue why, but the Dora results are better on average to me, so I train them anyways
>>24944 they've been cooking consistently for months so it's probably coreweave being shit again https://wandb.ai/novelai pretty sure they're working on their chatbot model atm
>>24946 >pretty sure they're working on their chatbot model atm Yeah I think so. Don't really care for textgen that much tbh and I've been using Claude Sonnet for a few days for shits and giggles, it's more than good enough. I just want V4 already.
>>24947 >I just want V4 already. dunno I wouldn't expect it any earlier than october
>>24945 Latest version of lycoris way slower in comparison with anything below 2.2.0.dev7 on 4090 cards, asked a few guys to test it and they have the same shit. Old locon configs trains slower. Tried to train dora too and can't say its way more better than locon for me, might done something wrong but it feels even worse, have you done anything with learning rates or just use the old params and it turns out to be better?
>>24948 >dunno I wouldn't expect it any earlier than october Pain. But it makes sense.
>>24948 wouldn't surprise me, they're basically uncontested in the field of nsfw imagegen for people who are scared of command lines so it makes business sense to fight where there's ground to be won i'm just not sure there's much to win in chatshit because that market already exists - there's people paying for shitty services because they have no taste and then people with taste abusing gray area methods to get top end services for free. how does a, probably, 13b subscription model fit into that? overlap with their imagegen subscriber base to keep them around when they get bored with imagegen, i guess. i think the turk is just an add zoomer retard who goes chasing after exciting whitepapers like a dog.
>>24951 >i think the turk is just an add zoomer retard who goes chasing after exciting whitepapers like a dog. bro literally went mask off at the end lmao what's wrong with providing more services at the same price? nai's textgen side is arguably and ironically its weakest service now and no new users sign up for it, it's all about the imagegen you don't complain if the butcher throws in a free sausage or two when you buy steaks, right?
>>24952 you write like a bitch
>>24949 I use similar settings to my normal lora with dora, all of my configs should actually be in my mega drive, I upload them alongside the model itself
>>24953 i know what you are but what am i
>>24955 you're not actually disagreeing with me, you're just thinking about it from a customer's perspective and i'm thinking about it from nai's perspective as a business
>>24956 i am disagreeing with you
BOARD OWNER-SAN, IF YOU ARE ALIVE: would you be ok with me making a textgen thread here? the ones over on 4chan are exploding due to some recent drama so it would be nice to have an obscure spot where people can fall back to when they want to stay on topic. if you'd like /hdg/ to stay /hdg/ only, that's cool.
Did anyone make a 1=2 Lora yet?
>>24959 You talking about b-lora? I haven't tried it but considering it targets specific blocks, 4 and 5 to be exact, it probably has no use outside of base SDXL Maybe still worth testing, but I'm personally not interested in it
>>24960 Cool.
>>24960 SEXO
>>24961 No, I'm talking about the artist 1=2, aka oet, aka 1 equals 2 www.pixiv.net/member.php?id=3068113
>>24958 speaking of text, runpod is currently running an imagegen contest on their discord. theme is "celebration", reward for submitting anything is $25 credit which will get you 15-20 hours of big boy text models or like 50+ of smaller ones. requires: discord, runpod account, two minutes of effort.
>>24967 DAME DA NE DAME YO DAME NA NO YO
>>24958 Wasnt the rule just "dont post anything ilegal"? It's most likely fine
>>24966 Hot, love the tail on the second one
Is it possible to get Nahida's pupils done correctly on AutismMix?
(3.01 MB 1920x1280 00141-3021859462.png)

(3.07 MB 1920x1280 00061-2519162277.png)

(3.71 MB 1920x1920 00120-4172542868.png)

(3.99 MB 1920x1920 00169-3221792748.png)

Easter bunny girls >>24971 I should gen some cat girls too >>24975 These look cute, nice style
>>24976 Thanks, and hot Easter bunnies!
>>24975 Very cute >>24976 Yeah you should
(3.10 MB 1664x2288 catbox_chwniz.png)

(2.87 MB 1664x2288 catbox_57g9ii.png)

>>24979 cunnymiya...
>>24981 NOW WE'RE TALKING
>>24981 >>24983 Are these NAI through the webui extension btw?
>>24984 nvm just checked, they are
>>24981 >>24985 I'm in awe at how well your prompts work despite being this fucking messy, it's a small miracle
>>24986 Thanks, it's because I fold my prompts 1000 times through the power of autism
>>24986 How are they messy? The only weird thing I see is "expressive eyes, beautiful detailed eyes" part which looks like something I'd see prompted during anythingv3 days
>>24988 You can just remove that, I think it does help a little if you don't use adetailer, but with adetailer faces and eyes are fine anyway and you can free up the prompt space for other stuff. As for why it helps make better eyes, it helps probably for the same reason {{{green eyes, looking at viewer}}} helps. If the AI pays more attention to the eyes, it does better, and mentioning them more makes it pay more attention. But in all honesty, these prompts are bad and messy because they are month old prompts that I recycled in png viewer. If I'm feeling lazy, I just browse through my gens and pick any image I thought looked good as my base prompt and modify it as little as possible to not fuck with whatever voodoo magic made it work good. NAI is insanely sensitive to prompt order and also to where 75 token chunks end so even small changes can fuck everything up. Uh anyway, have some nagatoros
>>24988 >How are they messy? They look like they were written by someone shitfaced drunk and under heavy dentist-tier anesthesia BUT THEY WORK. WELL.
>>24990 Well, the first is certainly true...
>>24989 >As for why it helps make better eyes, it helps probably for the same reason {{{green eyes, looking at viewer}}} helps. If the AI pays more attention to the eyes, it does better, and mentioning them more makes it pay more attention. Yeah, but I don't really think "beautiful eyes" and "detailed eyes" really do anything useful, back on local model it changed eye style, with NAI I think it does nothing, really. I think putting more emphasis on green eyes would work better (but really just inpainting works best) >Uh anyway, have some nagatoros Nice 'toros. >>24990 >written by someone shitfaced drunk and under heavy dentist-tier anesthesia Sounds like a great condition for proompting
>>24991 Godspeed you magnificent drunken bastard. >>24992 >Sounds like a great condition for proompting Probably is.
What artist combos do you guys suggest for AutismMix? If i'm mostly genning loli or that stuff
(2.22 MB 1288x1632 00062-1200839260.png)

(2.27 MB 1288x1632 00049-2774430555.png)

(2.36 MB 1288x1632 00051-1145465641.png)

(2.16 MB 1288x1632 00059-1343165374.png)

>Found a Chloroform lora that works My degeneracy is growing every day
>>24996 now do one where she's into it
>>24997 how can you be into someone drugging you unconscious
>>24998 (blush, heavy breathing, heart pupils:1.2)
>>24996 find a timestop one and reenact those touhou doujins or no balls
>>25000 https://files.catbox.moe/mg83p2.safetensors Here's the LORA I got it from civitai but they took down the lora because that site is dogshit
Is there any way to figure out the activation tags for loras I forgot about or loras deleted from civitai?
>>25003 Look up metadata. If the meta its purged, owari da, you can only guess
>>25002 Nakadashi them all
>>25006 >>25007 Uoooooh Kuro... my greatest weakness...
bird women............
top tier but damn i wish that foot was better
>>25012 meant for >>25011 obv
>>25014 last one overall
no age slider lora mannnnnnnnnnn fUCKKKKKK
>>25016 {{{loli}}}, {{{aged down}}}, {{{petite}}}
>>25018 (at this point i'm convinced we won't get actual controlnet shit for v3)
>>25018 >>25019 Cool shit but what I really want is openpose and anime lineart controlnets. Probably never getting those properly for XL.
>>25020 >but what I really want is openpose and anime lineart controlnets Same. All I need is openpose. >Probably never getting those properly for XL. Probably not during V3's lifespan at this point, V4 will probably come out first. Did base XL get decent controlnet tools or is it still stuck with the original shit ones?
>>25021 Sorry, not tools, models.
>>25021 > Did base XL get decent controlnet tools or is it still stuck with the original shit ones? Don't know really, I only used txt2img for local XL. I have this link bookmarked https://huggingface.co/lllyasviel/sd_control_collection/tree/main But afaik those models have issues, some of them basically don't work.
>>25023 >But afaik those models have issues, some of them basically don't work. Yeah that's what I was referring to Controlnet seemed to work just fine with the 1.5 models, I wonder why the XL ones suck so much.
>>25023 https://github.com/lllyasviel/stable-diffusion-webui-forge/wiki/ControlNet-model-download Also had this bookmarked in case someone wants to try >>25024 don't know but tbh with XL models txt2img is much more satisfactory, controlnet doesn't feel as important. Still would be certainly be nice to have
>>25025 >with XL models txt2img is much more satisfactory I agree but again I just want openpose to force specific angles and poses.
>>25023 SEXO
but what did he mean by this
>>25024 i'm starting to think XL is just a shitty model >>25029 natural language interpreter, maybe. the NAI crew have planned on the eventual total integration of all of their services from the start.
>>25030 >natural language interpreter, maybe fucking hope not, i despise boomerprompting even if it's an okay bandaid for sloppy tagging or entirely missing tags
>>25031 the goal would be to let users select a paragraph of text and generate a decently matching picture. it's very hit and miss right now. multi-vibe transfer looks like a step in that direction - character vibes, setting vibes, etc.
>>25032 >the goal would be to let users select a paragraph of text and generate a decently matching picture so shit nobody actually cares about
>>25034 so. shit. NOBODY. actually. cares. about.
>>25035 have a stroke about it if you want
>>25036 go back
>>25037 i've been a better poster than you for longer
>>25038 >i've been a better poster literally who?
>>25041 sexo
>>25043 nice
>>25043 Stealing your mix again. Again.
>>25046 s-s-she's fine, right?
>>25046 very nice. love the last one. moar
>>25046 Loss! Loss! Loss!
I know you're here. Post the chicken.
>>25053 Thank you. I'm off to getting banned for the watermelon.
haven't been around much, real life hit me like a wall if I flew off a motorcycle at 100mph+ Has there been any new things with the state of local? Base models, tools, and the such? For reference, I was gone when Dall-E and NAI2/3 crept up.
>>25055 Howdy, yeah local has been getting some new stuff. Some ponyfag released a SDXL model and that's the base for what most of us are using right now: https://civitai.com/models/257749/pony-diffusion-v6-xl The actual model I'm using right now is AutismMix SDXL: https://civitai.com/models/288584?modelVersionId=324619 For tools, we got Stable Diffusion WebUI Forge. Which gave us shorter gen times, especially with SDXL models: https://github.com/lllyasviel/stable-diffusion-webui-forge
>>25056 >pony model fuck, but I guess it's something Any info on it's training parameters and content? >Forge is it a reskinned WebUI or is actually an improved product?
>>25057 The backend is apparently different but it uses the same UI. It is way faster though.
>>25057 improved, illyasviel is the madman that brought us controlnet
>>25057 it's a massive speed-up for GPUlets and a moderate one for high-end GPUs. Overall it's pretty great, but there can be rare issues with extension compatibility (loractl for example). One concern is that it hasn't been updated for a month but I don't think Illyasviel abandoned it, just busy with smth else for the time being.
(2.38 MB 1248x1728 00023-1812477950.png)

(2.47 MB 1248x1728 00029-1315693275.png)

>>25057 >Any info on it's training parameters and content? No idea about training params. As the model name implies, a majority of furry/pony stuff. I don't know the ratios, but there's still a considerable amount of anime in there. Unfortunately the trainer did dumb shit like: >hashing artist names and characters to 3 character combinations >implementing a retarded score system I do not recommend using it as is because of the schizo default prompts and the default style, use the mix that anon linked in >>25056. Personally I prefer the autism confetti mix. It's still a horsefucker/furry model at base, but it's the best we have at the moment. It actually works well for anime to the point where most of the time you won't remember it's a horsefucker model. Don't worry about seeing random furry/pony shit in gens, you won't get them unless you specifically prompt for it. Inclusion of those in the dataset unironically helped it understand H/body anatomy better. Pros: >Higher base resolution means hands and fine details are overall better >It has better token comprehension and the model is more consistent at respecting most/all of your prompts >more responsive to prompts and hence character loras trained on it are a lot more flexible (clothing parts, poses) >It's the only good model for H we have Cons: >default style is a point of contention, you'll see a lot of thick lines and hints of sepia everywhere due to majority amount of Western art. >aforementioned dumbass decisions by trainer >It's the only good model for H we have
>>25057 >and content Supposedly an even split between furry, pony, anime and western comics. The troon who made it poisoned it in more ways than simply mixing sources (like hashing tokens out of spite or just for personal use and the overwhelming amount of scat and homemade fleshlights) and it suffers from a snapping issue, probably from the lack of data.
>>25065 >snapping issue What do you mean by this?
>>25066 It "snaps" to certain data in the dataset, mostly when it comes to poses
Quite pleased with this one.
>>25068 very nice
(3.36 MB 2080x1920 catbox_slhahg.png)

(2.72 MB 2080x1552 catbox_c54ge5.png)

(3.38 MB 2080x1920 catbox_0318hm.png)

(3.30 MB 2080x1920 catbox_fxgnp0.png)

>>24768 >>24847 thanks for the loras, I like the combination of them.
>>25070 glad you like them! (am the tyakomes guy) i've been putting way too much time into making a ton of lora (dora actually) so feel free to take another look at my mega occasionally, the SDXL stuff across characters and styles is up to, I think 40 or so separate lora just as a heads up, which will serve as a note for everybody, my styles and characters are not trained with the score tags, so they might be a bit weaker then some others if used with the score tags. and i'll just repost the link here for now for brevity sake: https://mega.nz/folder/OMY2RRxQ#zYaokRZV361WSNImfrBCrQ
>>25071 What are the epi_zero and anime_in_real_life loras?
>>25072 epi_zero the is the name of an [artist](https://gelbooru.com/index.php?page=post&s=list&tags=epi_zero+), so it's a normal artist lora, the anime_in_real_life lora is an attempt to do exactly that, anime characters in a real life backgrond, it's fairly limited in capacity, and it's not amazing, but it does seem to improve backgrounds either way.
>>25073 I see, thanks.
>>25071 Damn man, when is Illyasviel going to wake up? Just radio silence, seems like gonna have to do a separate git clone just to use dora.
You've bought her game, right?
>>25077 Touhou 8?
>>25077 damn those are some good wings... sex with misty...
>>25075 np, hope you enjoy using them! >>25076 I am pretty sure they aren't going to come back to working on forge tbh. I've been using the catboxanon fork since it came out, and it works without any issues, I know that auto has merged DoRA support into main now, so there's that. From what i've seen, lllya has been working on a private repo recently, maybe a new paper, maybe some other thing, not sure what, but they clearly aren't interested in maintaining forge. I'm a firm believer that DoRA is better then base LoRA though, so it's likely I won't be baking LoRA versions of my models anymore, too much time spent testing a model I consider strictly worse then the other version I baked, you know?
>>25080 She's very cute
>>25081 >I am pretty sure they aren't going to come back to working on forge tbh. I really hope this is not the case. Can't go back to auto and fuck the noodle soup. >but they clearly aren't interested in maintaining forge Maybe he's just gonna do it sporadically. Not ideal but better than nothing.
>>25083 I honestly hope so as well, the benefits of forge over auto is far greater then people think. But i'm going to be honest, if forge falls behind, i'm just gonna have to get used to using the noodles, despite my brain absolutely despising anything node based.
>>25088 nice, those were the closest I got after maybe an hour. I really want to try and get the "arms at sides" version but that might require some img2img
>>25089 >but that might require some img2img What if you do arm up > scribble the dick > inpaint > inpaint the arm down? anyway, slightly scuffed
>>25092 slight disney face but otherwise really nice
>>25091 That's a...pretty uncanny mix of clay stop-motion models and pastel aesthetics. Reminds me of Mrs Tweedy from Chicken Run.
>>25096 Now this is great, I gotta gen some dark scenes as well.
>>25098 cool, thanks
Sex with old women
im still using 1.5 based models because im a free thinker and i dont want to retrain my loras
>>25103 based 1girl standing enjoyer
>>25103 based, anything other than 1girl/pov sex isn't worth genning and i'm not shitposting
>>25103 Based, I do miss the 1.5 backgrounds.
I remember back when we used to post here solely because /hdg/ was unusable during the cumrag wars. Now it's twenty times worse thanks to all the schizoshitposting and (((niggerspam))). Can't wait for Iran to turn isntreal into wasneverreal.
>>25107 >I remember back when we used to post here solely Time sure flies and yes "main /hdg/" is fucking aids right now.
>>25109 Kill yehudi scum? Don't have to tell me twice!
so is someone gonna make the text thread or what
>>25111 Is there really enough posters to justify making a new thread? Feels like it's like 8 dudes left here now lol.
>>25112 probably not, i'm just a little bummed out by the lack of traffic but don't have much to post myself
>>25111 I'm just waiting for it tbh, keeping up with the free proxies among all the /aicg/ trooncord drama is exhausting.
>>25113 I hope to post more, but get caught up in trying to create storylines between pictures instead of just spamming gens
>>25114 fair, and i've barely touched chatbots so i wouldn't have much to contribute personally >>25115 i keep ending up on pseudo-renders which don't really fit the thread
>>25113 place needs new posters but no real idea where to get them from. not risking posting the link on current 4cuck hdg in any way for sure
>>25117 the /d/ thread seems a lot less mentally ill, ironically
>>25118 /e/ as well. they probably didn't get the textfag shitposters and schizos that came along with naiv3 release. at least that's hypothesis why /h/ became so bad.
>>25101 Is it just me or is this style really fucking fragile and picky? It breaks depending on the character/copyright (which must be placed right after the mix otherwise they don't work and if I place them before it's almost guaranteed to fuck up the style) and it more or less requires me to prompt the copyright otherwise it almost ignores the character (though with some characters this happens regardless)
>>25119 havent been in /h/ for a hot minute, what happened? Last i remember people were getting nae nae'd for posting flat chested characters and there was a surge of people posting cringe racial play stuff?
>>25119 oh, /e/'s got tons of lolifags. i have no idea how i'd advertise there without being rude though, seems like a pleasant place.
>>25122 edit the metadata and put the link in there as usual? "if you want to cunnypost without having to catbox all the time check out the metadata" or something like that
>>25122 I personally wouldn't mind seeing more people in this thread, I need to post more myself too, but I just spend all my time making lora, not genning, so it's kind of hard to find good gens to post
>>25120 It could have something to do with me switching out dino (dinoartforame) with quasarcake. Overall it's been an improvement but it's also a lot stronger
>>25122 Drop breadcrumbs Don't be open about it Cunnyfags are used to working hard for what they want to see
>>25121 >people were getting nae nae'd for posting flat chested characters and there was a surge of people posting cringe racial play stuff? there it's just endless shitposting for any reason last few months. nai vs local, loli, racial play, comfy vs forge, you name it, it's a /v/ parody at this point. latest one was just relentless interracial spam and trolling, while people getting banned for small breasts and/or slightly loli-ish faces, some schizo screaming "pedo" at everyone (like literally ctrl+f "pedo" gives 49 hits in the latest thread). I honestly still mostly visit that place simply out of habit, it's really bad
(2.54 MB 1920x1280 00016-3032015395.png)

4cuck refugee here, freshly getting the fuck out of there. been testing out perturbed attention guidance. i have been liking the results so far
>>25123 my idea was to post a loli gen with stealth png attached and just a base64-encoded link there, just like that. I wanted to do that on hdg but decided against that eventually >>25128 did you have this place bookmarked before?
Hilarious, got a 3day for posting nagatoro. Not that bans mean anything, but just the principle of the thing is aggravating considering that HALF OF THE THREADS in /h/ are for franchises like pokemon where EVERY CHARACTER is canonically fucking 10, and the trannitors are getting their panties in a twist over nagatoro.
>>25128 Welcome. >>25130 I got banned for a Hikari gen.
>>25125 When it works it works really fucking well but like I've said it just breaks with some characters (or it turns characters that work perfectly into a gacha) It's weird, I've never seen a mix do this
>>25128 uooh best* girl *shared with roll.exe
>collarbone
I luv edits (and Kuro and Miyu cunny) >>25130 rough.
>>25132 >>25125 Testing it again, really need to figure out how to keep the mix working with all characters (surprisingly enough it can fuck up Cirno) I switched back to your negs and it seems to be more consistent style-wise, my negs (human focus preset but manually copypasted and tweaked) are probably too restrictive
>>25135 Thanks for reminding, only tried that before several times. Gotta add pantyshots to anime screencaps now.
>>25136 Think I understand how to wrangle it a bit more but not fully (also NAI is shitting the bed with the gen times and the bugged outputs). 1: Higher CFG (tested with cfg 10, think you can go above 10 by typing it in, check the baka on the left even if it's enhanced and the base gen reimu) can really elevate the mix (and it can reduce the issue right below). Enhancing (right reimu) still fucks it up a bit but still, fucking nice. 2: Copyright is a must but even then it's not a guaranteed fix depending on where you put it, it can still hallucinate OCs even with overrepresented 2hus. Need to insert it as late as you can get away with to avoid breaking/overriding the mix, it's a balancing act. Need to experiment with the character tag more but ditto from above except it works best overall when put right after the mix (and not before or it might override the style) 3: Lolis can be hit or miss, sometimes you get really nice faces like >>25092 and sometimes you get porcelain doll or disney-esque faces with eyes that are WAY too big - petite and loli in pos/neg respectively or a combination of the two can mitigate it but I'd rather not do that and just roll/inpaint tbh, "petite" doesn't work as a good substitute for loli with this mix. Thanks for coming to my TED schizotalk.
>>25138 Ah, yes, Inubashiri Momiji. Riiiiiiight.
>>25138 I really like this style, I have NAI would I be able to make a style LORA If I wanted to say, use it to make local gens in this style? I should be able to right? considering theres style LORAs for actual AI artists on patreon
>>25140 You probably can if you cherrypick the gens with the exact style you want and they're at least 1024x
Anyone here has experience with wildcards? Ive been working on a setup that generates characters with the most important parts randomized (hair, eyes, body type, etc) along with a lewd/sex act and a background. So far it works but ive come to realize that sex acts rely too much on the correct tags to work so now im trying to come up with a different setup that generates the character depending on the sex act (no paizuri if the character has a flat chest, for example)
>>25137 BASED. Truly the way AI was meant to be used. Hopefully this catches on more on /a/
>>25137 based
(2.52 MB 2184x3200 image(3062c).jpg)

I'm so retarded, I searched the list of boards but didn't actually try to just add /hdg. Thanks for helping the retard-kun whoever posted the image with base64 metadata.
(1.94 MB 1632x1440 00171-2238966360.png)

>>25145 No problem.
>>25145 Very nice, welcome. Hopefully we won't get schizos here from the posted link.
>>25145 Welcome, this place is kinda slow now but I try to post when I find time to gen.
>>25147 Entire splitoff thread got nuked so I'd say it's unlikely at this point
(2.18 MB 2176x3200 image(2566c).jpg)

(2.88 MB 2176x3200 image(3189c).jpg)

(2.28 MB 1920x2688 image(3190c).jpg)

(1.98 MB 1920x2688 image(3191c).jpg)

Some stuff I wanted to finish making for a while but couldn't be bothered because there was nowhere to post. Don't expect 4 pictures every time, but I'll probably try to make at least 1 a day to keep the thread a bit busier. >>25149 I think even the blue haired LMAO poster didn't crack the code, even though the message was addressed to him as well, so unlikely.
I'm fucking done with /h/ jesus, place is cancer getting good advice is over there isn't worth it for all the bullshit.
>>25151 /h/'s moderation is comically bad, I blame trannitors for everything that happened to that place
(2.01 MB 1248x1728 00007-2551908533.png)

>>25150 >blue haired poster Shame, I liked his gens. Looks like he caught a ban though because his posts got nuked before the thread did.
>>25155 He gets banned every day I think, he'll be fine. Although he did say he will stop posting now too, so a shame he didn't find this board indeed.
I hate people that upload a lora to civitai then rip off the metadata from all their sample pics and give zero instructions on how to use it god fuck offfffffffffff
>>25150 Very nice, I like your pics. Could use a catbox for 2nd one for sure.
>>25152 all the boards have extreme moderation either overbearing like /a/ and /vr/ or barely any like /h/
>>25159 > or barely any like /h/ unless you post cunny then hope you love getting ISP banned
>>25160 most likely thats due to reports and not natural moderation like some boards have
>>25159 >>25160 Had a look at the current thread and the mods delete/ban pretty much everything except for the low quality spam. And yes cunny always gets deleted first asap, especially lately.
>>25161 Other boards dont have ISP bans At best you get a block on the /v/'s for reported though I remember this one time I got all my posts deleted for posting fully clothed lolis on /vg/
>>25158 https://files.catbox.moe/me3t8k.png I genned this base resolution with normal pony, it gives way more variety than autism I think. Plus sometimes you can get very colorful images like >>25145 Upscaled with autism confetti though, more stable for limbs and many other things. >>25162 It's actually so weird to me that they actively encourage these low quality posts. I guess they want /hdg/ dead? No clue what's their endgame.
>>25164 jannies quickly grow to hate the boards that they janny and malicious compliance with the rules is how they cope
>>25164 Thanks. Cool style mixing, didn't expect to see no loras used
>>25166 I'm just coping with how bad loras work out for me usually, kek. Trying to get the maximum out of pony with whatever artists are actually there, because $25 monthly for nai is not my cup of tea. Only occasionally use fkey atm, and before I posted some madoka on /h/ using shirow. Otherwise just revisiting base unmixed pony with innate artists.
>>25167 how did the pony artist tag decoding project turn out?
>>25168 I think there are plenty of styles found now, but I only used vum, which was one of the first very strong ones found. Didn't have time to go through the whole table picking out good stuff yet... But if I get tired of mixing unhashed ones I'll probably have to.
>>25168 >>25169 Any good tokens that work with autismmix
>>25150 Good stuff >>25153 uoooh
>>25159 >/vr/ Wow you weren't kidding I went in /vr/ and made a thread asking when the PS3/360/Wii will become retro and that shit got caught fast and I was warned, last time I was on /vr/ the threads moved slow as shit.
(2.84 MB 1664x2288 catbox_aywc04.png)

from the kasuga dora I baked for pony the other day https://mega.nz/folder/OMY2RRxQ#zYaokRZV361WSNImfrBCrQ/folder/6chGzABZ
>>25172 i got a 6 month ban or something a year ago which i evaded for making a thread about the best retro youtubers to watch because i wanted recommendations i would have learned my lesson with a 3 day
>>25172 I have bad news, they're already adults and old enough to drink (in Europe)
>warned for saying "dyke hairstyle" (low quality post apparently) >banned for a cotton field joke >adult vulpo woman removed >groidspam intact There's enough human-sized turdspam in /h/ to rival your average jeet train derailment.
(3.66 MB 2176x3200 image(2751c).jpg)

(3.45 MB 3328x1984 image(2903c).jpg)

(1.61 MB 2176x3200 image(3221c).jpg)

(2.91 MB 2176x3200 image(3222c).jpg)

Experiments with Takamichi v2. Very strong lora, rather difficult to cleanly upscale. Unsure if worth the effort, looks a bit weird? >>25176 Why is some guy always sperging out when you post on /h/?! Also can I request miyu_(swimsuit)_(blue_archive) with your style?
>>25178 >Why is some guy always sperging out when you post on /h/?! No idea, I've noticed it's at least two people now. (((They))) start to sperg the moment they see a style that isn't the groided Patreon mix, whenever it's not a groid or when you post the same character twice in a row (the noface miqote samestyle spammer is mysteriously exempt from these "rules", weird coincidence huh?) >Also can I request miyu_(swimsuit)_(blue_archive) with your style? Sure, here's the armpit but feel free to request anything, I'm too lazy to change my prompt on mobile if I'm prompting for myself. Also it took an infuriatingly long time to get something even remotely decent, either NAI doesn't like her or the seed generator is shitting itself as usual. Probably both.
(2.12 MB 1440x1440 00061-1902430128.png)

(2.66 MB 1440x1440 00004-2220390506.png)

(2.12 MB 1440x1440 00019-2481834159.png)

(3.62 MB 1920x1920 00000-3552035239.png)

Posting some random gens to back into the swing of things
>>25125 By the way, the mix looks really nice at very low cfg too (as low as 3 if you don't mind some slightly wonky eyes and/or inaccuracies from time to time, 3.5/4 is a good balance), seems like the cfg slider is more like a strength slider when it comes to this particular mix
>>25180 sexo
>>25184 >Looks like it would be good with traditional_media I agree, the paper texture or some light millipen lines would look pretty good, lemme see if I can introduce either without destroying the style. I think I'll start using it at cfg ~4 more often, I like how soft it looks. >or maybe upscaled with higher cfg You can do that? Do you just modify the cfg while on the enhancing UI?
>>25179 Thanks for pits! Well if it really doesn't wanna go well, it's all good, otherwise something more zoomed out like this would be great: https://archiveofsins.com/data/h/image/1712/67/1712678289708329.png Idk I just love her swimsuit. Pony knows her, but not the swimsuit version.
>>25184 Oops, this message didn't load for me before I posted mine. Those are pretty good.
>>25186 Catbox the image, archiveofsins doesn't allow hotlinking >>25187 Yeah his look good, think the seed generator finally stopped shitting itself, I was getting really inconsistent stuff earlier (and the damn bunny pin? kept getting ultra melted)
>>25188 Here, I think it's also your gen: https://files.catbox.moe/1b1h7h.png archiveofsin loads for me if I just press enter in the address bar.
>>25185 You can change whatever you want when you enhance but I usually just change the sampler to euler. You could change the prompt entirely if you wanted to, same as regular img2img
>>25189 Yep that's also mine. I should gen more Ruka now that I figured out how to wrangle NAI into making her semi-accurately. >>25190 I actually didn't know that, I just assumed it used the same settings and prompt as the base gen for convenience.
>>25189 >archiveofsin loads for me if I just press enter in the address bar. uhhh
>>25192 Hmm never tried on the phone before but it works as well for me. You get this message, go to address bar, press edit url, press enter. On PC it's easier to just copy-paste the link instead of clicking it, but you can also just click, then press enter in the address bar.
>>25193 That's odd, if I tap the link directly I get that message (and it won't work even if I refresh/tap the url and press enter) but if I copy the link from your post and paste it THEN it works
>prompt (((tengus))) >strings and pompoms are inconsistent as fuck, don't feel like inpainting them >string and pom pom (clothes) in negs, tokin hat in pos >doesn't work REEEEEEEE
>>25195 Do I seriously have to gen them without the hat at all and then inpaint it?
>>25197 Thanks, very nice, print looks very authentic too. Great things one can do when the model just knows everything and isn't pruned by some degenerate horsefucker...
>>25178 might finally be time for me to switch back to local
>>25195 >>25196 >tokin hat, hat, string, pom pom (clothes) in negs >hat won't come off >none of them work hmmmmmmmmmmmmmmmmmmm
>>25202 Can you try Momiji?
>>25199 It's kind of pain in the ass to get to this "high fidelity" though, can sometimes spend hours on 1 pic (although I try not to these days). >>25200 >Oops I forgot to remove bocchi the rock kek
>>25205 Well, fuck. Skill issue it is.
>>25206 >no headwear >didn't think about it Actual skill issue holy shit
>>25202 >scientists say this is the last thing you see before you die Glad I tried your prompt out, holy shit
>>25205 Does the mix ever do this to you? This is like the third time, fucking jumpscared
>>25209 Sometimes but definitely not as often as it would without rescale
full circle to my first childhood crush
(2.40 MB 2176x3200 image(3234c).jpg)

(1.78 MB 2176x3200 image(2200c).jpg)

(1.39 MB 1500x2432 image(2740c).jpg)

(3.10 MB 3200x2176 image(2231c).jpg)

My Taiga's still up! Although it's usually the US timezone jannie that gives me 3-days while I sleep... Adding takamichi v1 lolis to this post, think I catboxed all of them to /h/ before, but let them be here too.
>>25215 You should really do zoomed out more often, tits look great. But also nice tongue.
>>25215 Damn, she'll devour the dick.....
Getting a decent masturbation gen turned out to be more frustrating than I expected and I still feel like it doesn't look right here.
>>25213 Maybe even the troon janitor couldn't resist the cute Taiga cunny.
(790.11 KB 832x1144 catbox_r7ubt3.png)

(733.98 KB 832x1144 catbox_d18lpy.png)

(789.36 KB 832x1144 catbox_uf8ney.png)

(858.46 KB 832x1144 catbox_z1kehk.png)

not able to upscale right now, but I was able to gen some solid stuff with the shun(ny)lora i'm working on right now
>>25213 first is i think the nicest looking pussy i've seen from imagegen
>>25223 inpainting model
(249.10 KB 250x250 gi_lick.gif)

>>25223 Polka pits....
>>25213 Saw your post, anon. What settings did you use to merge the models? It's been ages since I merged one myself, but it didn't sound like you were using block weights.
>>25218 seems fine to me
>>25227 Nope, no block weights just adding difference between autism and pony in the good old checkpoint merger tab.
>>25229 sdxl model merging rapes RAM hard btw
>>25220 funny how people back in 1.5 days thought that upside-down faces and poses were literally impossible for ai ever, kek.
>>25229 Thanks. Was confused for a moment why you'd set the multiplier to 1, but I guess it makes sense for add difference
>>25232 You can go down the "math" rabbit hole and say that since js2prony is already a mix of pony and something else, it contains say 80% pony, so we need to set multiplier to 0.8 to replace exactly that amount of pony with a mixed autism version. BUT first, you don't know how much base pony actually remains in js2prony, and second, 1.0 just werks (for me).
(2.17 MB 1248x1728 00184-491505401_cleanup.png)

(2.27 MB 1536x1536 00187-2933784439.png)

(2.09 MB 1536x1536 00189-3732995089_cleanup.png)

(The batteries are dead)
>>25234 >>25235 kono... ###
>>25235 I can't believe hoshino would do this
>>25237 from the looks of it, she didn't know she would be doing it either. probably a request from sensei
(2.35 MB 1664x2288 catbox_eqybp2.png)

Does anyone still have the borvar lora around? Tried generating with novelAI but not the effect i wanted
(2.56 MB 1224x1824 catbox_9f4nqb.png)

(2.58 MB 1224x1824 catbox_kt4t2y.png)

(2.53 MB 1224x1880 catbox_ykub1e.png)

(2.40 MB 1224x1880 catbox_hhrupp.png)

multiple views is quite good at coherence these days. I feel like with a lora on comic pages and/or controlnet help for the layouts and general scale, you could go quite far. Maybe not temporal coherence, but at least different sex acts and clothing.
kinda worried about v4 now tbh
>>25245 I'm still holding out hope for the cascade finetune but it's going to take a miracle for them to reach their donation goal. Even then I still doubt we would have anything near a NAI equivalent model within 2 months
>>25246 there's no real rush anyway, if v4 has no artist tags it's completely doa unless they implement actual ipadapter inster of their shitty semi-broken "ViBe tRaNsFeR" and even then it'd be a shitty bandaid solution if it's good then yeah, cascade finetune is gonna be a doomed but given what happened today....
>>25245 they probably want to see how people react to lack of artist tags in furry and if "just vibe bro" cope works out. I don't think it will
>>25248 >they probably want to see how people react to lack of artist tags in furry and if "just vibe bro" cope works out. I don't think it will 90% of the dicksword lapped it right up and i very much doubt they have any reason to look anywhere else for feedback i really hope they will but i already know they'll just grow complacent thanks to the bootlickers
>>25248 to add to >>25249 aside from "just vibe bro" from both analtan and the community one of the most prevalent copes was "i don't even use artist tags on v3"/"styles are not that important"/etc
>>25249 aren't those discords circlejerks anyway? like could they actually call bullshit out?
>>25251 >aren't those discords circlejerks anyway it's the official analtan dicksword, it's THE circlejerk so why would they look anywhere else for feedback after trying to pull this kind of thing?
>>25252 >so why would they look anywhere else for feedback after trying to pull this kind of thing? because they're actually interested in profits, not some zoomer trannies rimming them? no idea if it's gonna massively affect their revenue though, I hope it does lol
>>25253 >because they're actually interested in profits, not some zoomer trannies rimming them? i'd hope so but then why try to pull this shit? it makes no sense >>25253 >no idea if it's gonna massively affect their revenue though, I hope it does lol it'd fucking better, probably the only way to make them realize this is a terrible idea. what's probably gonna happen is that furfags will sub for a month or two, realize the gap between it and local furshit isn't as gigantic as the anime one and go back to local. analtan will see this spike, commit to pruning artist tags at first then pikachu face as soon as they see the downward trend
>>25254 >then why try to pull this shit? probably trying to balance risk and profit. risk of twitter mobs coming for them, card companies, potential future legislations against styles, etc, they're likely worried about all this and are considering if they can get away without artist tags. no one has mentioned that yet but I wonder how the "cub" (or what the fuck furry loli is called kek) is doing on the new model, could they have pruned that as well? I feel like considering all the payment processors bullshit lately, cunny being pruned for future models is a possibility too
>>25255 >twitter mobs frankly irrelevant even if furfag artists are much more aggressive when it comes to le holy artoids >card companies paddle exists for a reason >future legislation against styles bit too late for that already innit? even if they come up with AND manage to pass something like that it's gonna be unenforceable, you can't copyright a style or we wouldn't have 20 CSRB ripoffs >cub works but it got reworded as young on e6 iirc, loli also works
>>25256 seems like they just left out artist tags which doesn't fit with the whole nsfw/payment processor fearmongering otherwise they'd have neutered it (get it?) it much harder
sex with feet
So cascade guys accepted my dataset augmentation for brikot (replaced some gelbooru images with higher quality ones and added some missing ones). Also planning to give them replacement images for gemba (he has half his stuff censored with giant black bars on gelbooru). Don't know if anyone's interested and you won't make it in time making anything from scratch, but if anyone already has an artist dataset of images that are missing/bad quality/censored on gelbooru, you can still send it to them until maybe tomorrow? After that they start caching stuff so probably will be impossible to change anything.
(2.49 MB 1280x1920 00081-1813719831.png)

(2.84 MB 1280x1920 00092-3073682066.png)

(2.73 MB 1280x1920 00152-2951944097.png)

(2.57 MB 1280x1920 00292-3210919724.png)

>Cheers!
>>25260 sexo >>25261 based
I wish we could do text already, I want to see what it would come up with
>>25261 >>25263 is this just inpainting on the girl?
(376.03 KB 1024x1024 file.png)

>>25264 Yeah >>25263 was also inpainting the guy's outfit and some of the text
>>25261 What the fuck he's literally me
(1.64 MB 2176x3200 image(3281c).jpg)

(1.78 MB 2176x3200 image(3319с).jpg)

(1.82 MB 2176x3200 image(3332c).jpg)

>>25268 >spend all that time genning and inpainting and then convert it to a bitcrushed jpg with hood spongebob levels of crust cmon man we have reasonable filesize limits here
(5.73 MB 2176x3200 image(3281c).png)

>>25269 You are very wrong. I do jpgs with lowest compression, there's no visible difference with .png, I just double checked, can't even see a pixel difference in A/B comparison at 100% zoom. It's just because I genned this on base pony. The other 2 I upscaled with 0.5+ denoise on autism, so artifacts disappeared, the first one I didn't have much luck. But here, I attached a .png for you, you can check yourself.
>>25270 Ok, I take it back, I clearly see difference on this pic, kek. It's just so fucked up from pony that I didn't realize it got even more fucked up.
(6.36 MB 2176x3200 image(3332c).png)

(5.21 MB 2176x3200 image(3319сс).png)

Really though it's augmenting pony's artifacts. Even if I take the darkest gen from here >>25178, which is supposed to be affected the most by bit crushing, it looks good (it's genned with autism confetti from the start), but the current artist combination together with base pony is really doing a number on quality of the images even before any .jpg is introduced. It's probably because the innate artist had a lot of low quality jpgs in the dataset.
>>25272 why engage in this kind of autism when you can just use pngs to begin with if it's too big just use tinypng or superpng/pnggauntlet
>>25273 When the scenes aren't dark pngs are 11-12 mb for my resolution. I mean on this website it's not an issue, on all other sites it is.
luv edits
>>25270 >>25272 >it's not shitty compression, it's a machine accurately aping shitty compression what a future lmao
>>25274 so just run them through superpng or pnggauntlet or similar you don't output directly in jpg do you?
>>25272 the jpegging on base pony really is completely fucked. tried using it and it's just way too frustrating
>>25278 Doesn't help even for a 9mb png... Also it took 8 whole minutes for one image!
>>25277 >>25279 Well you get who I replied too, sorry!
>>25279 something's not right, it should save quite a lot more
>>25281 It's set to optipng level 7 (highest). I mean I can introduce png palette but then it will be a lossy compression again so I don't think it's worth it. Look at my images though: I currently tried the first one from here >>25178 and tell me it's going to be compressed with any success? I understand flat styles, but that's not usually what I go for.
>>25282 Are you just using img2img for your highres upscales like >>25178 here?
>>25283 Sometimes img2img, sometimes multidiffusion, whichever works best and spawns less atrocities for a particular image. Although as you can see I fucked up and the first image is a mess (left leg, thigh).
>>25284 I see. Mixture of diffusers has been acting kinda weird for me, sometimes it works really well even with ~0.45 denoise, sometimes it starts consistently spawning garbage even at 0.25 (euler a). Like this pic got all covered with fingers when I tried 2x upscaling, disgusting shit lol.
>local actually useful for something huh. Couldn't get NAI to do BJ face. Still had to use it to fix some retarded errors (some of which still visible since I cba)
>>25286 :>= didn't work? probably could with some artist tag
(8.95 MB 3072x2304 image(2810c).png)

(7.86 MB 3200x1984 image(2902с).png)

(5.66 MB 1856x3200 image(2939c).png)

Muh dokas (in glorious .png!) >>25276 Damn black bars. Kinda wish I could see more, you know?? >>25285 Wait this pic is local? Or are you just upscaling NAI with pony? Try only using horizontal or vertical splits, just do one side exactly equal to your upscale size, and divide the other side in 2 or 3.
>>25288 >Wait this pic is local? Or are you just upscaling NAI with pony? That one is fully local. https://files.catbox.moe/vobe56.png Lack of darks in pony annoys me a bit so I tried pushing it with vectorscope and some loras https://civitai.com/models/352581?modelVersionId=426797 https://github.com/Haoming02/sd-webui-vectorscope-cc >Try only using horizontal or vertical splits, just do one side exactly equal to your upscale size, and divide the other side in 2 or 3. Interesting, I'll try it. Usually just set the pre-upscale pic size as the tile size.
>>25289 Oh I actually used the dramatic lighting already, but it felt like it nuked half of pony (in terms of camera angles and poses), but I probably used it at too high strength. I thought vectorscope is weaker than init noise, but I have it too, thanks. This pic of yours is currently my favorite out of at least the last 2 /e/ threads by the way!
>>25288 i admire your persistence in trying to make shirow masamune get along with other artists lord knows i made a lot of fucked up greasy snake torsos trying on 1.5
>>25291 Thanks, I actually found it pretty easy to work with for Madoka on pony, but... the fact that I have to use base pony to get the best images out of it makes me wish anime girls didn't have hands.
>>25292 And I mean hands are bad on base pony but easy to fix upscaling with autism later. But with shirow... fuck!
>>25287 yeah I didn't try for artist tags since I needed OG style. Might've been a case of NAI needing more space aswell or smth. I have a hard time inpainting solid blushes aswell, usually have to use local to force them
>>25262 thanks >>25267 cute gens, love the style. Care to share the mix? >Stupid sexy long ears Gonna post some more with her soon. Any requests?
>>25290 >Oh I actually used the dramatic lighting already, but it felt like it nuked half of pony (in terms of camera angles and poses), but I probably used it at too high strength. Yeah it's pretty overfit, but not impossible to use. A proper noise offset lora would be really handy though. >I thought vectorscope is weaker than init noise, but I have it too, thanks It's pretty strong if you activate alt. noise, also allows correcting the colors a bit to push the gens away from the meme sepia. >This pic of yours is currently my favorite out of at least the last 2 /e/ threads by the way! Thanks, cute b&w girls too! Nice to see that /e/ janny is not fucked in the head like the one from /h/ and allows small chested girls. Maybe I could try sneakily posting some Kuro gens...
>>25286 >Couldn't get NAI to do BJ face What >>25287 said, vacuum fellatio gets aliased to :>= on danbooru Admittedly I'm a promptlet and I'm prompting in the background so I still have some trouble with it but I'm sure there's a way to get it consistently
(1.80 MB 1280x1920 catbox_8pxhpu.png)

(2.17 MB 1280x1920 catbox_hd44n3.png)

>>25295 Have some catboxes. >Any requests? Partially submerged.
>want to gen yordles >need to play the fucking STYLE gacha >most styles are absolute garbage wow, i fucking hate roaches
>>25261 how do you get these to work reliably, i keep getting generic girls
Cascade's got some ways to go but it's already looking promising
>>25301 dangerously close to pony/ef long torso territory but fingers crossed
(1.40 MB 896x1152 s-750252536282646_00001_.png)

(1.22 MB 896x1152 s-750252536282632_00001_.png)

(1.09 MB 896x1152 s-750252536282608_00003_.png)

>>25302 Maybe wasn't the best example since I was just happy to get it working. But I'm surpised it got things like fingernail color and holding sign plus the style mix is a little basic but it's there
Here's a basic comfyui setup for anyone wanting to try out cascade. Just basic prompting with rescale added, laid out to fit on half a 1440p monitor https://files.catbox.moe/ht7aja.json https://rentry.org/resonancemodel https://huggingface.co/stabilityai/stable-cascade/tree/main You will need the following files: ComfyUI\models\clip\cascade\reso_proto_delta_e5_te.safetensors ComfyUI\models\unet\cascade\stage_b\stage_b_lite_bf16.safetensors (other versions should still work) ComfyUI\models\unet\cascade\stage_c\reso_proto_epsilon-e5.safetensors ComfyUI\models\vae\cascade\stage_a.safetensors (already included)
>>25303 Looks a bit rough for now but tbh I already feel like I prefer its style to pony somehow.
>>25303 It's like maybe 20% done?
>>25304 Is it possible to do hires fix upscale right now? Was trying to but it throws errors on the sampling stage after image resizing with upscaler model, maybe I've used wrong nodes for cascade or something
>>25307 It seems you can but I haven't been able to get it to work yet. Here's a workflow that has it that I'm currently trying to wrap my head around https://comfyworkflows.com/workflows/2e72682e-dc58-4ceb-8dab-49ddfd074292
>>25308 Well nevermind I can't get this one to work either
Haven't tried comfy, but img2img works for me in capgui. (forgot to set denoise so it was the default 0.75, but it actually turned out not too fucked kek)
(2.81 MB 1280x1920 00084-3044082809.png)

(2.83 MB 1280x1920 00076-1981900865.png)

(2.71 MB 1280x1920 00091-2761296579.png)

(2.71 MB 1280x1920 00081-2769371821.png)

>>25298 >Partially submerged. yes watersports is spoilered
(1.82 MB 1472x1472 cq6g9m.png)

(2.54 MB 1472x1472 xzlkej.png)

(1.64 MB 1536x1024 3q2qm6.png)

(2.21 MB 1472x1472 ado4pv.png)

>>25314 based
>>25314 >>25316 Tanya sexo. Wonder if you can get her closer to the manga look though, anime flopped hard in terms of sexo
>>25317 What do you mean, bird girls are very popular these days, including ducks. More seriously, the artist for the manga and the official artist for the LN don't have a whole lot on danbooru so NAI isn't going to be able to do that well directly, but I tried. Using ONLY the official artist tags was really bad, so I kept my other mix but lowered the weight. Ignore the somewhat mangled hands.
(1.39 MB 832x1216 tmpzrdp_em0.png)

One more that worked well, removing some of the stronger face influences.
>he comes here when they don't give him (You)s
>>25319 there was a tanya manga lora for 1.5, otherwise seems its all the same duckface'y. Thanks for trying
>>25318 uoooooh blonde old men
(5.59 MB 1536x2304 Gen_00242_c.png)

(5.48 MB 1536x2304 Gen_00249_c.png)

(3.65 MB 1536x2304 Gen_00253_c.png)

(2.12 MB 1024x1536 Gen_00137_.png)

Didn't expect to have much fun with the prototype because it's crude and underbaked and spawns horrendous shit, AND YET I'm having lots of fun.... I guess many here already had this experience with naiv3, but it's very new for me to just go to gelbooru and pick some artists and slap them together. And styles are really deep and profound here, it's no pony, let me tell you. I think if this had inpainting I would just switch over fully right now. Kinda hard to fix every single error by just upscaling. Last pic is raw without upscaling and color-correction, the other 3 I put a bit of work in.
>>25324 sovl
>>25324 And here I thought the deep frying was part of the prototype, but apparently the settings of whoever made all those other example images were ass.
(2.54 MB 1344x1728 catbox_7fkjss.png)

(2.40 MB 1344x1728 catbox_edpcmh.png)

(2.29 MB 1344x1728 catbox_s8m3bm.png)

(2.90 MB 1344x1728 catbox_34q7cs.png)

>>25326 nta but most 4chins barely know comfyui let alone a working workflow for stable cascade
>>25327 the last two but especially the last one look surprisingly pretty decent
Where are you guys getting the prototype?
(1.15 MB 896x1152 catbox_hr7cd2.png)

>>25329 See here >>25304 This catbox has a more recent basic comfui workflow. The catboxes I posted above are using the smea extension at a higher base resolution. You could also add hires fix instead, this is a good reference https://openart.ai/workflows/data_lt/stable-cascade-img2img/lWT2VR6mFDO9cvBqEsxV
(9.48 MB 2304x3456 Gen_00341_c.png)

If anyone wants to experiment with styles on cascade, I made this .csv from original anime.csv. It includes only artists, leftmost column is original name to be able to search for artist on gelbooru, rightmost column is formatted for cascade "by artist \(no underscores\)" https://files.catbox.moe/9mxuo9.csv
>>25332 looking quite promising.
These images look much better. Whoever was posting the initial cascade images on 4ch wasn't doing them any favours.
>>25334 always the case, remember the first week of v3
I know its pretty useless right now, but any way to train cascade lora from those prototype weights? Really curious if that could be any good. Already tried onetrainer and there is no way to change the base model, or I don't know how to do it without errors. Jordach scripts doesn't support loras for now, so I don't know what else to try.
>>25337 sexo
(7.56 MB 2048x3072 Gen_00557_c.png)

(8.28 MB 2048x3072 Gen_00575_c.png)

>>25334 I could probably help that poor guy in his cascade shilling enterprise, but I can't make myself gen anything /h/ appropriate. All I'm genning is lolis. I keep trying to type medium breasts and my fingers refuse to move. Send help.
>>25339 Yeah tbh that could help a lot, the pics posted here are certainly better than fried eyes shit that I've seen on 4chan. Some people are already getting disappointed with the whole cascade idea due to poor examples.
>>25339 Yeah that's why I never post my gens If I could make /h/ appropriate stuff then I would have made a pixiv and/or some ko-fi shit but the only thing i'm into is cute small girls or degenerate stuff so everything I do never makes it out of my PC
fizrot's back as moemortem and finally doing full H btw
(8.06 MB 2048x3072 Gen_00677_c.png)

>>25340 Probably won't make a difference. Methinks whoever hates pony enough and has spare money already donated. Just looks like anons are poorfags. But I can post this paizuri (and a catbox to lolis) in response to that shill guy if he comes again next thread. Is that what the normals like!? I can't tell, I heard they like paizuri! I'm like Mark Zuckerberg here trying to speak to human race but all that comes out are noises of some alien-robot hybrid. >>25336 You tried looking through civit loras for base cascade? For example one guy posted his settings in comments here: https://civitai.com/models/314077/animegao-kigurumi-sc
>>25344 Well it wasn't too bad, 100+ posts discussing cascade and... lolis. Anons even forgot to post images for a while in the heat of all the polemics. Classic /hdg/.
>>25343 I'm so glad he's back even if the "just for fun" alias means we probably won't get stuff as good as before (though it's still really, really good)
>>25343 >fizrot's back
>>25347 hey that's my fedi joke
>>25348 your what
>>25344 >Probably won't make a difference. Methinks whoever hates pony enough and has spare money already donated. Just looks like anons are poorfags. Most anons just don't want to donate for any reason, sure, but some just want to see actual results before considering anything. Too many grifters in the space
(2.18 MB 1248x1728 00050-1071758412.png)

(2.02 MB 1248x1728 00047-1650213005.png)

(2.33 MB 1248x1728 Sch.png)

(2.08 MB 1248x1728 00046-781825780.png)

>>25350 Unstable diffusion poisoned the well, for sure. All that money and zero results. Zero details, no epochs, no writeups, nothing.
>>25344 >For example one guy posted his settings in comments here He did mention sd-scripts, there is abandoned branch with it, yeah, so worth to try at least I think. Those gens looks surprisingly good though, is it pure cumfy? Mind share a flow for that?
>Don't even gen porn sometimes >Just cute girls in cute outfits to see how the AI designs them Man, I hope im not subconsciously like gay or something
>>25354 Less is more, well at times it is.
>>25352 whew, #4
>>25356 upper body and close-up can be pretty nice.
>>25352 well, basado.
(7.69 MB 2048x3072 Gen_01014_c.png)

(6.50 MB 2048x3072 Gen_01009_c.png)

(6.12 MB 2048x3072 Gen_01021_c.png)

>>25353 Oops, I wanted to answer to you yesterday and I just forgot...... My workflows are actually very simple, I took them from CAPGUI/workflows/init.py. Using basic Euler A too, nothing fancy. It's probably just an artist mix, a bit of gacha rolling for good seed and a bit more rolling for upscale seed. I did swap to dpm++ sde 6 steps instead of euler 12 steps for stage b only in i2i, but I'm not even sure it makes any difference. Here's a workflow with a discount wingless cirno prompt (it uses the same artist mix as these >>25339 images, I like it the most for now) t2i: https://files.catbox.moe/nt88ej.png i2i: https://files.catbox.moe/pa50r3.png
Best XL model?
(8.63 MB 2688x3456 Gen_00003_.png)

(7.93 MB 2688x3456 Gen_00007_.png)

(8.19 MB 2688x3456 Gen_00008_.png)

(8.26 MB 2000x2571 Gen_00009_.png)

>>25359 Thanks for sharing flow, nothing fancy as I can see. Tried to bake lora and for such an early undertrained base model it works fine, I'm even impressed by that consistent background with fine fence net on the last pic, but the girl not that great on that gen tho. Dont know why this artist was not included in the dataset, maybe because of prototype, but whatever. Clip encoder from your flow is also throwing errors, when using with lora, guess it's not updated to that. What is "repeat latent batch" or "modelsamplingstablecascade" with shift nodes for?
>>25362 I use that clip encoder because with default one tag weights below 1.0 do not work at all, like (butterchalk:0.5) and (butterchalk:1.0) produce the same result, making mixing styles properly impossible. I heard that prompt control extension also supports comfy++ weights, but I haven't tried it. If you want to use my adv clip encoder node setup, you will need to do a hacky think for it to support cascade: custom_nodes\ComfyUI_ADV_CLIP_emb\adv_encode.py Change line 275 to if True: And restart comfy, then it will work. >What is "repeat latent batch" or "modelsamplingstablecascade" with shift nodes for? Repeat latent batch is just batch size. Shift is useless, CAPGUI says you can create seed variations with it, but is sucks for that purpose, so you can just reroute past it and delete it I think. It's pretty cool that you managed to make a lora btw.
(22.75 KB 1782x244 notepad++_B0CYrxmdaQ.png)

>>25364 >And restart comfy, then it will work. Nope, still no luck with that to work. Assuming, that I'm correctly moved python intention in this part, when embs_l, _ completely removed the custom node just wont load on the start >It's pretty cool that you managed to make a lora btw. Here, test it too, if you wanna https://litter.catbox.moe/rhzeoe.safetensors takes up even more space on the same rank than for the XL
>>25365 WTF no idea what I was smoking when I typed that. It's line 245. It should already have an if that checks if it's SDXL. We just wanna say fuck it, always use that branch.
>>25366 Yea, now works fine
(6.72 MB 2048x3072 Gen_01048_c.png)

(7.37 MB 2048x3072 Gen_01103_c.png)

(6.62 MB 2048x3072 Gen_01198_c.png)

(7.76 MB 2048x3072 Gen_01203_c.png)

Kinda flooded /e/ with cascade images. That board is too damn slow. Time to take a break tbqh desu senpai famalam. Want to test too many things. >>25361 Sex-crazed woman. >>25354 Hey I like cute designs too. You should post more of them here!
>>25368 >That board is too damn slow It is, and /e/ is a redundant board, but it's nice to have a place almost free of schizoposting and the bbc janny.
(3.17 MB 1664x2288 catbox_qognia.png)

(3.08 MB 1664x2288 catbox_0ypgag.png)

(2.93 MB 1664x2288 catbox_0z4fx3.png)

(2.96 MB 1664x2288 catbox_kaiff5.png)

>>25371 very cute
>>25371 >crotch cutout AND white pantyhose t-too lewd!
>>25374 Hnng, nai is the best at lolis. Great pubes on the first one. artist:focke wulf <- love this guy, unfortunately he looks like ass on cascade even though plenty of images went into the prototype dataset. Makes me sad.
>>25375 he's great. his style is heavy on fine detail so cascade will probably improve on it as the model converges. >>25376 all that hopium about NAI releasing their older image models as they iterated went down the drain pretty fast when NAI stopped iterating lol sorry if i posted any of these before, finally going through the downloads folder. i bet you can guess what picture i used as a vibe transfer for #4.
>>25377 He's too great on detail, which makes me think resizing him to 1024 bucket size is a waste. He needs some close up crops of his 4k images in the dataset as well, but I guess it's solvable with lora.
(2.64 MB 1664x2288 catbox_hlnhmo.png)

(2.55 MB 1664x2288 catbox_bjsaql.png)

(2.58 MB 1664x2288 catbox_s527ag.png)

(2.36 MB 1664x2288 catbox_04f05r.png)

>>25380 calling the yakuza
>>25381 >>25381 I like the idea of the yakuza taking a plane to USA or wherever out of japan to gun down some dude making horse anime girl porn, lmao
>>25381 >>25382 I don't get it
>>25383 japs didn't draw porn of the horses the way you'd expect with its huge popularity because the IP owners made it strenuously clear they didn't want any made, the joke became that the yakuza was involved because horse gambling
>>25383 In order to use the names of real racehorses, the Umamusume franchise had to get permission from their owners. Because of how intertwined horse racing is with gambling, there's rumored to be tons of yakuza involvement in horse racing. Unlike other franchises, Cygames has repeatedly asked people not to make horsegirl lewds and has actually taken down some. People meme that this is because they are afraid of the yakuza. This has further led to the meme that if you lewd the horsegirls the yaks will come for you personally.
(3.14 MB 1664x2288 catbox_jyj2bn.png)

(2.76 MB 1664x2288 catbox_geyu4l.png)

(2.91 MB 1664x2288 catbox_dj3z36.png)

(862.57 KB 768x1024 20240324032556-2472005452.png)

(870.90 KB 768x1024 20240324032810-691016376.png)

What even is cascade? A new model? Why does it only work on comfy and what exactly makes it good?
>>25387 https://stability.ai/news/introducing-stable-cascade One of stability's models, one of the stronger points is supposed to be very fast trainability compared to XL. One of the furries trained a prototype model (anime+furry) based on this arch and it showed pretty passable results considering the relatively small dataset used, and he's going for a full finetune, claiming no censorship and full artist tags. There seem to be some hiccups with financing the whole endeavor though.
>>25388 I see. An uncensored model would be neat.
(1.06 MB 768x1152 00006-348218930.png)

(821.64 KB 768x1152 00079-4237447357.png)

(1005.73 KB 768x1152 00093-1921499028.png)

(949.48 KB 768x1152 00282-2396394581.png)

I've been trying to find the best way i can take this really unique style i've made in SD1.5, and somehow make it replicable in XL. Hearing that training a LORA with only 8 gigs of vram is basically impossible/takes too long, and i also have no idea were to find that ipadapter model for XL to do ""instant loras"". It seems my only option really is to just take these pics (and others, no idea how many i should gen and pick to make a good style lora) and make the LORA, I have experience making SD1.5 loras, specifically of celebrities, but that's it. Any/All help would be hugely appreciated. Resources used: LORAS: https://civitai.com/models/12523/pc-98-galgame-elf-retro-semi-pixel-art-style https://civitai.com/models/19416?modelVersionId=23046 https://civitai.com/models/82098?modelVersionId=87153 NEGS, Which seem to be how 60% of this style is even possible i noticed, removing them, not having the combination of all three, completely changes the result. https://civitai.com/models/163204/aissist-negative-embedding -Both V2 and Supercharge Also upscaled with 4x ultrasharp, i use SD forge.
(707.28 KB 768x1152 00302-1157038189.png)

(999.51 KB 768x1152 00306-1022622171.png)

(1.06 MB 768x1152 00309-3601777926.png)

(1.10 MB 768x1024 00313-817343964.png)

>>25391 Thanks for this rec. This tool has tons of potential but i am awful at Pony. It's too esoteric for me to get a solid grip on, but i can see the echoes of the style im looking for in my first try in pony. First pics are in a different pony model, the last one is AutismMix, which i think is the one i should be using with StyleComponent.
(1.24 MB 1866x936 hhhhhmmmmmmm.png)

>>25392 thought i attached a screenshot to this, but its not showing up.. but heres what im working with. If some Pony experts wanna tardwrangle me into figuring this out that'd be rad. Style Component may very well work.
>>25388 >supposed to be very fast trainability compared to X I dont know what magic jordach did for training, but with kohya sd-scripts its even slower than XL, at least version with big parameter count
(9.86 MB 2176x3200 image(3406).png)

(8.76 MB 2176x3200 image(3427c).png)

Took break from Cascade. Testing Olympus XL v2. Cons: -can't do difficult poses, can't do difficult angles, 700 flying fingers per hand way too often -tried to make a girl masturbate: "female masturbation"? didn't work. "hand on chest, hand in panties" (effortlessly used here >>25268), "chest" sometimes works, "in panties" - no chance. "hand on crotch/own crotch" at least?! Okay this works 1 time out of 5 but the fingers are flying in all directions... -too scared to try to prompt for sex -everything besides cleft of venus is a body horror type of pussy (sometimes even cleft of venus produces huge chunks of meat where it supposedly wanted to draw clitoris) despite using one of the artists this model was finetuned on (curated dataset!) who mostly draws pussy close ups....... -trained on mature women artists mostly, was able to neg it though, negged it hard! -high denoise upscale doesn't make image cleaner like with pony, instead depending on artist there's more grain/color bending/artifacts/chromatic aberration -eyes too stylized making them messy even at high resolutions, couldn't be bothered to clean them up TL;DR: it's an animagine finetune Pros: -pictures look pretty Kinda torn on this one ngl. The pros are quite alluring.
>>25395 The pics are indeed pretty.
>>25395 cute
(905.45 KB 768x1024 00315-2908405860.png)

(940.20 KB 768x1024 00017-3052703624.png)

(840.93 KB 768x1024 00055-185749251.png)

(1.02 MB 960x1280 00075-3196609194.png)

After many hours of testing, and having to learn how Pony works to still mixed results, Style Components XL is pretty interesting. It has a lot of potential though it's a shame it lacks in being able to pick up facial styles and stuff. My progress in going from monstrous abominations to something resembling a consistent style and quality has been, to say the least, painful. Also thought XL was supposed to be way better than 1.5 on prompt adherence, Why is it somehow as stupid as the 1.5 model i'm used to? Still can't handle "holding glass of milk in one hand", will randomly add clothes to a nude character, LORA adherence is all over the place which is already defeating the purpose of why i'm upgrading to XL. 'Course it could be the LORA, but its inconsistent between models. Some models work better than others but the ones i'm seeing suggested most are all over the place. Anyway hope this isnt considered spammy, can't really get help for this anywhere else and after just asking once i've already gotten more help here than anywhere else. Thanks.
>>25393 I've found some success cropping the majority of the images to square 1:1 ratios around the face. I couldn't be assed to figure out how the code actually determines and applies the crop to make it automatic with nudenet. maybe some other time
>>25381 ive actually had a couple horse images that weren't even explicit taken down
>>25399 some other time indeed. burned out hard trying to figure this out so the idea's on ice for the foreseeable future.
ded place. wish /b/ wasn't so shit for cunny
>>25402 Meh, even cuckchan threads are pretty dead. Genning got stale once again.
>>25403 even worse they're not just dead, its mostly spam/arguing/boring gens gotta give you guys credit, at least this thread's nice in every way. >and helpful
>>25404 those threads peaked a year ago
(3.57 MB 1664x2288 catbox_bgoah5.png)

(3.53 MB 1664x2288 catbox_vww90y.png)

(3.08 MB 1664x2288 catbox_0y6qmr.png)

(3.32 MB 1664x2288 catbox_9i7zxg.png)

>>25402 I don't usually gen that much (i'm usually training instead), but I happened to be genning a bit today, so I guess i'll post to give the board a bit of life too.
>>25407 Very cute
(3.64 MB 2432x1664 catbox_3bg8hy.png)

(3.19 MB 1664x2288 catbox_tln197.png)

(3.05 MB 1664x2288 catbox_g3q8iz.png)

(2.39 MB 1664x2288 catbox_rnkkc3.png)

>>25408 Thanks! It's been a while since I genned any of hifumi, i'm sure some of the board veterans know of her. and of course suzuran is too damn cute not to gen. (most of these gens are test gens, the first one was playing around with forge couple for a small bit earlier today)
>>25406 ironic so did /lmg/. yet the tech is getting better.
>>25410 guess excessive ai coom consumption really does fuck with your brain
(906.16 KB 2150x3035 shupo_0001.jpg)

>>25409 >forge couple I have to give that a try again as pic related made wanna gen something similar
>>25412 made me wanna gen something similar* oof
>>25412 god damn i love white pantyhose
>>25412 damn shupogaki! reminds me that I need to update my lora for them now that they have a ton more art
>>25414 Yeah, me too. >>25415 Based
>>25416 I'll make sure to post when I get around to it
>>25418 try sticking no pupils in your negative as well
(1.48 MB 2184x1496 00041-4189130937.png)

aint no way the 4chan thread has real posters now has to be all bots
when are we getting a model that's easier to prompt on than pony? everytime I try I just bounce back cause it looks like fucking shit and seems to default to the same few poses. Apparently people use specific loras to unfuck it now or modify the score shit with more?
>>25402 site's just a place to talk about things if you want more people here then you have to let them know it exists first
(1.02 MB 768x1152 00000-1570668286.png)

(1.12 MB 768x1152 00001-1570668286.png)

>>25422 XL in general just seems like a nightmare to prompt for. I can never get even half as good results as i do in 1.5 makes switching less justifiable when picrel is so nice and fixable with more experience.
(2.55 MB 1728x1344 2.png)

(2.27 MB 1728x1344 1.png)

Forge couple seems significantly less finicky than previous regional prompter extensions both in general and when it comes to lora separation. It'll mostly do what I tell it to do with minimal wrangling as long as it's not too complex. But that might also just be pony prompt comprehension helping, I've not tried it with 1.5 models.
>>25425 >fixable make if you heavily inpaint in NAI. 1.5 looks like shit with hands and hair, poor understanding etc. Not worth it either
>>25430 Vidya getting in the way
(3.21 MB 1816x1816 00016-404899015.png)

>>25430 This stupid fucking bitch, 145 rolls worth of solid and liquid rocks GONE in five minutes. (still got her)
>>25435 is now a good time to tell you I got her on the first 10 pull? If it wasn't free I wouldn't have pulled for her either, because, i'm not gonna lie, I actually really don't like her that much. I wouldn't have made a lora for her if it wasn't a request from a friend
>>25436 Nah I'm not salty, I've had incredibly consistent sub-30 luck from Ideal City up until Typhon's banner. This was the first rough one but luck is a skill :^)
>>25437 fair enough, my luck is usually really damn shit, but despite only using free pulls, I've managed to get 2 of both banner units
>>25438 >my luck is usually really damn shit I have shit IRL luck and incredible gacha luck, I think the absolute """worst""" was Ideal City itself at 200+ rolls for Pozy, I had maxed out Doc Croc 2 (my beloved) at like <100 rolls.
>>25439 compared to me spending literally all of my pulls, some 700 or so trying to max pot eyja alter, I only got her to 5
>>25440 hi derrian
>>25441 of course, only you would know that.
>>25442 kek How's it going?
>>25443 so-so, been feeling really burned out recently, life has been fine for the most part, but eh just getting tired of being the one people blame when they can't train a good lora lel
(3.34 MB 1664x2288 catbox_vj7oyw.png)

(3.27 MB 1664x2288 catbox_kiifnx.png)

(3.18 MB 1664x2288 catbox_46hkz7.png)

(3.30 MB 1664x2288 catbox_qga9gj.png)

>>25444 might as well actually post something instead of just text, here is some more hifumi in various outfits (and styles)
>>25444 >just getting tired of being the one people blame when they can't train a good lora lel Yeah I've noticed that >been feeling really burned out recently Burnout hasn't set in yet but I've been cycling the same prompts and characters over and over
>>25446 yeah, it's not so much that I'm burned out on AI art specifically, but rather the people. because, like, I am still enjoying genning a fair bit, in fact, today is probably among the most i've genned outside of testing lora in a long while. And I still love making lora too, i've just been soured at the idea of providing my knowledge to everybody. like, I'm supposed to be taking a break from training lora, but I still trained 2 because fuck it why not. I'm good at it, and I enjoy it, so, if I'm making something for myself (finally. tbh) why does it matter if it's what i'm supposed to be taking a break from.
(2.26 MB 1440x1440 00021-1910194001.png)

(2.22 MB 1440x1440 00034-1031035439.png)

(1.91 MB 1248x1728 00103-3583865782.png)

>>25444 What sort of things are you taking the heat for, training settings?
>>25448 sexo
(1.64 MB 960x1440 00071-1159239941.png)

(1.64 MB 960x1440 00089-3993574253.png)

>>25450 >sexo yes I'm trying to gen less and post more
i am going to buy a 3090 and train a lot of loras that everyone will like and use and a bunch of anonymous internet retards will respect me now that i got the dopamine charge out of the way by imagining that it's already true, i'm going to wait until i can get one for like $600
(5.85 MB 2048x3072 Gen_02108_.png)

(6.48 MB 2048x3072 Gen_02104_.png)

(6.91 MB 2048x3072 Gen_02098_.png)

(6.65 MB 2048x3072 Gen_02100_.png)

>>25434 >>25452 Alternatively just keep genning in this style, I love it. Also I notice the number of {{{}}} on loli tag keeps increasing, I'm worried for you, kek. Could you try Eris Greyrat? Curious how it will look with your mix, though it will probably butcher her face. No need to spend anals on me, base res is fine with nai anyways. Also if nai needs help nailing details it's "eris greyrat, red hair, long hair, red eyes, black hairband".
(6.36 MB 2048x3072 Gen_01943_.png)

(5.36 MB 2048x3072 Gen_01945_.png)

(6.10 MB 2048x3072 Gen_01946_.png)

(5.37 MB 2048x3072 Gen_01956_.png)

Btw I'm not even using a loli tag. Cascade is just loli by default I guess. Or it learned the character well enough idk.
(6.20 MB 2048x3072 Gen_01981_.png)

Last one didn't fit.
Nevermind, I'm using a (loli:0.8) tag, I'm just sleep deprived and blind.
>>25442 Any plans on adding masked directories of datasets to easyscripts? Or am I just blind and can't find where is it?
(1.89 MB 1024x1536 Gen_02419_.png)

(3.53 MB 1536x2304 Gen_02412_.png)

(5.87 MB 2048x3072 Gen_02403_.png)

(3.51 MB 1536x2304 Gen_02433_.png)

Damn... "by abe yashitoshi" gives me some nai tier gens.
(4.10 MB 1536x2304 Gen_02480_.png)

(4.23 MB 1536x2304 Gen_02482_.png)

(2.20 MB 1024x1536 Gen_02467_.png)

"by kenket" According to boorus, the guy is a hardcore furry who hasn't drawn a single human female in his entire life. >>25458 "by abe yoshitoshi"*
(1.92 MB 1024x1536 Gen_02537_.png)

(3.62 MB 1536x2304 Gen_02593_.png)

(3.76 MB 1536x2304 Gen_02594_.png)

(3.71 MB 1536x2304 Gen_02596_.png)

(3.65 MB 1536x2304 Gen_02597_.png)

(3.69 MB 1536x2304 Gen_02599_.png)

Okay definitely enough now.
(3.79 MB 1920x1920 00176-3630299994.png)

(2.27 MB 1920x1280 00122-63287131.png)

>>25462 oh yeah bunny girls, I haven't genned that in ages.
>>25457 I *could* add them, I guess, I didn't think anybody would use it, so I wasn't really planning on supporting it
(3.61 MB 1536x2304 Gen_02650_.png)

(3.43 MB 1536x2304 Gen_02651_.png)

(3.32 MB 1664x2288 catbox_juj48p.png)

(3.12 MB 1664x2288 catbox_87bgfx.png)

>>25468 thats a really nice style
>>25468 god i want to plap this schizo retard
It's a shame NAI furry is mostly useless for me because of the lack of artist tags. >>25473 was the nicest-looking image I could get with it
>>25473 Base NAI can do an OK ruto if you aren't extremely picky about fin placement (which differs in every fan art and the actual canon fins are kinda ugly anyway).
>>25453 thanks, i really like that gangly look so i was pushing it i dropped my NAI sub, it's a lot harder to justify 25 bucks a month than it was $10 more than the 15 i was paying them for textgen anyway i'm hoping some localchad will train a good SDXL inpainting model because that's by far NAI's biggest advantage left
>>25465 Masks not very useful when you training style since you want information from the whole image most of the time, but its very useful for characters and concepts to improve accuracy and reduce the impact of backrounds on training. I've also noticed that loss schedule is only available when you select full huber loss, but it should also be enabled when you are using pseudo-huber loss aka smooth l1. Judging by this discussion could be a nice alternative for min snr gamma https://github.com/kohya-ss/sd-scripts/pull/1228/ >>25477 >i'm hoping some localchad will train a good SDXL inpainting model I haven't tried this one because I don't like cumfyui, but looks very promising https://github.com/nullquant/ComfyUI-BrushNet
>artist: tsurui, [artist:aaaa (quad-a)] Discovered a kino combo
>>25479 stupid sexy preggo bellies
(1.72 MB 1472x1472 00071-1612798256.png)

(1.91 MB 1472x1472 00063-2695875770.png)

(1.72 MB 1472x1472 00062-2765440312.png)

(1.82 MB 1472x1472 00034-4270367386.png)

>>25480 the magical source of sexo and happy feels
>>25482 no but I still bought an expensive express and half my table is filled with various flavored coffee
>>25483 my fellow caffeine addict
(1.17 MB 242x243 feet.gif)

>>25475 NTA but I need to experiment with it more, I kinda gave up after the first two days due to the lack of artist tags. If only Vibe Transfer was even remotely decent. I wonder if they had some special agreement with e621 (ie we'll let you scrape if you don't train with artist tags on) or if they're just covering their own asses because westoid furfags would pull them into lolsuits (which they'd lose because you can't copyright a style yet but still) left and right.
>>25487 What bothers me about analtan is their lack of transparency on the matter and just the SHEER unprofessionalism they exhibit when you try to get anything out of them or even give them any kind of feedback that isn't a bug report.
>>25468 >>25474 >>25475 very nice >>25488 they're a bunch of imageboard retards trying to build escapist heaven for themselves first and run a business second, set your expectations accordingly
>>25489 >they're a bunch of imageboard retards >set your expectations accordingly Yeah, I should've.
>>25487 I mean, you just know that if they had included furry artist names that the furry community would have gone absolutely apeshit. I don't blame them for wanting to avoid that but god damn it sucks.
Japan explicitly stated that using datasets to train AI doesn't violate copyright laws, so it's much safer to allow artist tags on a jap art dataset. Making same move with e621 would be riskier, they probably decided it's not worth the headache. I've seen claims that vibe transfer is really good, but I personally haven't been able to do anything great with it tbh, feels rather messy most of the time
>>25492 >Making same move with e621 would be riskier But why? Because of the constant lolsuits from furfags, because those lolsuits could give (((legislators))) ideas, because they're futureproofing their asses against some (not really but people probably believe it will be) retroactive law that will come out SOON™ or all of them combined? That's what I want to know. I hate not knowing shit.
>>25492 >I've seen claims that vibe transfer is really good, but I personally haven't been able to do anything great with it tbh, feels rather messy most of the time Either I don't know how to use it or it's shit, it feels like it's even worse on the furry model.
>>25493 until there's a serious, iron-clad US court ruling one way or another, everyone involved in generative AI is dancing under gathering storm clouds in an open field, hoping someone else will be the one to get struck by lightning first
>>25493 Any of those is a possibility, and they probably don't want to deal with any of this, even if the chances of this shit happening are slim. They pruned artist tags from v1, remember? That was before Japs ruled that training is fair game. v2 and v3 released after that. Yeah we don't know their reasons for sure but I think this is the most reasonable assumption. At least it's my headcanon for now.
>>25493 It's important to note that there are a significant amount of furries who are fucking insane as well as terminally online. It would be such a fucking clusterfuck.
>>25496 >Yeah we don't know their reasons for sure but I think this is the most reasonable assumptio (C)hoping that V4 will still have artist tags Still, shame about the furry one (not paying theboogie tho)
>>25478 hmm, I guess I can look into adding masking to the UI, but I'm not gonna be able to for a bit. as for the loss schedule, I thought it was for huber only, because all of the documentation pointed it to only work with huber
>>25478 ah fuck i don't want to touch this autism here we go
>>25500 >>25501 I'm not completely sure myself if that is correct, but I've tried to start training just from cmd passing smooth_l1 along with snr scheduler and huber_c 0.1 and it ended fine, I can't say it's too different from what I've already trained with just min snr, but it's fine too, haven't tested lora much, ymmv if you ever decide to try it probably. Besides that kohya_ss huber_c param states that it is also applicable to smooth l1, so
>>25503 >foot grab UOOOOOOOOOOOOOOOOOOH
(996.39 KB 896x1152 00068-109326439.png)

(1.07 MB 896x1152 00063-174420534.png)

>>25506 same but for artists cunny oc's
>>25507 And Alexander wept, seeing he had no more worlds to conquer.
>>25507 Also can you box the >shadman vault girl
(613.55 KB 896x1152 00082-1331059262.png)

>>25509 there is stealth pnginfo meta in images
>>25510 Ah thanks, didn't check. Is there an extension that works on this board like the one on 4chan?
>>25511 NTA but check OP
>>25510 > <lora:pony_colorful_lbw_5x:0.2>,<lora:lily_sdxl_v1:1> <lora:lilandy_sdxl_select2:0.43> Which loras are those? I remember encountering pony colorful before and couldn't find it.
>>25503 sexosexo >>25497 never enough porn for this series. Vignette needs another succubus doujin
>>25514 >never enough porn for this series. Vignette needs another succubus doujin s2 never so its forgotten anime now except for Satania memes Gonna have to gen doujins ourselves
>>25517 getting neets pregnant and watching them feel like shit in the mornings.. and then fucking them again anyway uoh
>>25507 >shadman YES
>>25518 Can't resist the neet angel cunny.
>>25521 sexo
>>25521 making biblically accurate angels
>>25523 >making biblically accurate angels Well...
>>25524 me in the middle so this is what gabriel geting mad would look like huh
>>25502 alright, I'll make sure to change the huber_c and scheduler args to be valid for smooth_l1 then
(8.12 MB 2400x3600 Gen_02879_.png)

(4.71 MB 1536x2304 Gen_02883_.png)

>>25477 Ah oops, you aren't subbed, sorry about requesting then, and thanks for genning anyway, saved! >>25497 >>25503 Nice.
i fucking hate comfyui man
>>25528 Why are you torturing yourself?..
>>25529 i want to try this shit >>25478
>>25527 Thanks! cute redhead >>25531 Thanks >should mix styles more often as it brought back some of my MOTIVATION certainly, mixing styles is fun
>>25533 uooooh
>>25532 touching isn't enough lazy brats #
(2.98 MB 1664x2288 catbox_wsfmo8.png)

(3.12 MB 1664x2288 catbox_sx0q8b.png)

(2.64 MB 1664x2288 catbox_7f9gkj.png)

(2.82 MB 1664x2288 catbox_yn0e82.png)

>>25526 judging by my results here, I guess I got masked loss working, I still gotta fix one other thing, then I'll push the update
>>25537 gladly
(3.54 MB 1664x2288 catbox_241nvu.png)

>>25536 https://github.com/derrian-distro/LoRA_Easy_Training_Scripts/tree/dev fixed the steps bug while I was at it, now it should calc steps in the same way as kohya proper. that was painful.
>>25540 I'm arachnophobic BUT
>>25536 >>25539 Nice, thanks. So, do you automate process to make those masks somehow? I found it's boring to make them via ps and created workflow for coomfy, but it's not very optimal, since it's not making full white on the character area and not very accurate either https://files.catbox.moe/qhzw1i.json Also found that the same hyperparams inappropriate when masking, have you changed them for your training? Not sure if things outside masks should be tagged or not either
>>25544 I'll let her hunt me.
>>25543 I did it manually via PS and a python script to remove their transparancy, I also used the mask I created to cut out only what I wanted to tag, and tagged the dataset according to that, all in all, I will never be doing that again
>>25546 oh, also, it works off of the red channel, so my masks are red and black. In my *one* test, the character was fully cooked by epoch 6, which is about 8 epochs lower than usual. My goal wasn't to learn how to use it, it was to make sure it worked. I still see very little reason to use masked loss tbh
>>25545 She will drain your wallet/bank account
>>25548 Jokes on her, she is now in debt
>>25549 damn..
What are some places to post AI apart from this one? Not necessarily loli, just anything. 4chan threads are now insufferable cancer, and this board, while nice, is pretty slow. I tried looking into some discords but the ones I was able to find feel like they're full of moralfag zoomers.
>>25551 I have no idea and I'm not in any image gen discords, I do share my gens with friends I have on discord.
>>25551 regrettably, current mass culture is cancer and the only two avenues to escape discovery and infection are racism and pedophilia, and the former not so much anymore the /e/ thread seems okay but there isn't much discussion
>>25553 Checking out the /e/ thread right now actually.
anyone know why autism is doing this using webui-forge set up the same way i used sd1.5 models
>>25555 >Denoising strength: 0.4, Hires upscale: 1.5, Hires upscaler: Latent (bicubic antialiased), Change the upscaler to something like Remacri. Latent outputs pictures like this with denoise below 0.5 and you shouldn't use it if you aren't sure what you're doing.
>>25556 oh. right. thanks
>>25551 I want to ATF (yeah three letter agency honeypot memes I know) and their LORA thread is in such a shit state they don't even know how to bake shit and the one guy who does know is shit at it, not only that but he's asking for money for said LORAs acting all smug about it lmao. They're just relying on CivitAI's uploaders before they get nuked.
>>25558 actually, they don't know how to train lora, I know exactly who you are talking about, and what I know for absolute certain is that they stole one of my tomls I keep in my mega drive and fucking butchered it
>>25546 >>25547 >fully cooked by epoch 6, which is about 8 epochs lower than usual That actually nice, but still have to spend more time preparing the dataset instead of just throwing the pictures in the oven. Oh, it's also probably could be useful to not learn annoying text/watermarks from images even when style training with enough autism curating dataset. Will try to bake with red/black later as well, but white seems to be ok too. Also tried schedule free optimizer and seems like it has absolutly no compute overhead with comparable results, unlike dora, in my quick tests it was like about range of 3-4x increase of lr's for adam version to bake properly, 10x completely fried
>>25560 In my comparisons with my normal settings (REX + CAME) the schedule free is just not worth using over it. For the styles, technically speaking, the only thing I would use the masks for is to mask off the text/signatures. I'm not sold on using masking for anything other than absolutely horrible datasets tbh
>>25561 >he only thing I would use the masks for is to mask off the text/signatures. I'm not sold on using masking for anything other than absolutely horrible datasets tbh True, me neither. Have you tried lora+? Any good?
>>25562 no I haven't tried it, and honestly, I probably never will, I was sold on dora
>>25563 Dora is fine, I like it either, but seems like minimal gains in comparison to a good locon. I would've train everything as dora, but it feels like simply doesn't worth to train 2 times longer for same thing, just trying to aim for "under 30-40 mins for one fine bake" config. Hope kohaku at least fix dropout, it's slowing down the training even more now on the latest versions of his package
>>25564 I haven't been finding any longer training times with DoRA vs not, 30-40 minutes for about 300-400 steps in both cases. probably a difference in config I guess
>>25565 This is your config? >>25173 Strange, I have similar batch, but not accumulation https://files.catbox.moe/w8l3s2.toml you have only caption dropout either I also asked a couple of guys with 4090 cards to test it and they said that they have worse performance too on the latest version, what is your gpu?
>>25566 a6000... yeah that could be a big reason why, I sometimes forget I have double the vram most do
>>25567 oh and yeah, that *is* one, but you should make sure to grab the most recently added one
>>25567 So you technically just running 3090 with double amount of vram, maybe 3090 is not affected or you haven't tried to compare other dropouts to notice that
>>25569 well, I haven't been using any of the dropouts because I usually only use network dropout, and DoRA doesn't support that
(2.34 MB 1200x1600 catbox_95mdc0.png)

(2.25 MB 1200x1600 catbox_yr85mg.png)

(2.03 MB 1200x1600 catbox_kn905d.png)

(2.42 MB 1200x1600 catbox_x571dx.png)

>>25570 I'm a little curious, how much vram does it consumes to train with bs2 natively, without checkpointing?
>>25572 about 42gb, it's just about 47gb if i'm training the TE as well, but I train Unet only
>>25573 Thats too much, I thought it's in the range of 30gbs for 6gb weights+updates at fp16/bf16 or something, even bs1 is just 18gb for unet only, dafuq, it's completely doubled
>>25574 yeah, it does, but the funny part is once you put checkpointing on, it's only 11gb, absolutely insane
>>25575 Yeah, it can go even lower than 8gb with fp8_base weights for absolute vramlets. Have any experience with deepspeed? I saw this https://github.com/kohya-ss/sd-scripts/pull/1101 and looks like something like 2x3090 should work now, at least on linux
>>25576 No, I've never used it because, well, I only have one gpu
>>25577 One day I will stop being a lazy ass, boot back to linux and try it, since I'm already running multi gpus for llm's. Probably when (if) sd3 will be available for train
>>25578 Yeah, that's valid, I'm just using windows because I don't like linux as a daily driver. Most of my focus right now isn't on lora training tbh, because I've got a larger model to help train.
>>25579 derrian man are you helping with cascade?
>>25580 no, i'm helping with WD
>>25581 They are cooking something? I thought they are just gave up after one of the last announcements
>>25582 yeah, we are. still too early stage to say anything concrete, except that we are training on cosXL instead of base SDXL
>>25583 >cosXL Thats based and nice to know. Fuck, if only stability just release a normal version from the start we wouldn't have to fix our gens with vectorscope/init-noise and other shit
>>25584 yeah, from what i've been told, some of the devs wanted to release a different version of SDXL much earlier, but safety basically said fuck you like usual
(3.04 MB 1664x2288 catbox_wac27o.png)

(3.66 MB 1664x2288 catbox_3ajtfx.png)

(2.40 MB 1664x2288 catbox_1wxvm7.png)

(2.59 MB 1664x2288 catbox_8lkbtz.png)

(835.74 KB 768x1152 00023-3738144960.png)

(946.26 KB 768x1152 00025-110232025.png)

(1.52 MB 1024x1024 00071-2751442918.png)

(1.52 MB 1024x1024 00073-627227019.png)

I've upgraded to XL, grinded 500 buzz on civitai recently and managed to make my desired PC98 LORA for XL. So far it's promising, but balancing settings catered to XL loras is fucking difficult as SHIT. some pointers would be really nice. gonna catbox a few so you can scrutinize my settings + prompting. https://catbox.moe/c/2p0t0a
>>25585 That is so overreacting, all this so-called "safety" just cripples the models
>>25588 yeah, that's why most of the core devs left. Of course, they haven't left AI as a whole, just SAI, and if things are going as I hope they do, we might just get a non fucked model from the devs of SAI... from another company
>>25587 the lora trainer for civitai is so drastically different that there isn't really any comparison to be made, you just have to figure out the settings yourself/ ask people that use the civitai trainer for pointers
>>25590 i think the lora from civitai turned out fine, its just my issues prompting for XL + the extreme fragility of LORAs i honestly just need help learning how to use this fucking model, its insane. pony looks like SHIT in 99% of my attempts to use it.
>>25591 yeah, I don't know how to prompt pony either, I just ignore all of the score tags, source tags, and the hashes, and I just let the lora I make carry the quality, really
(1.41 MB 1024x1024 00090-3767208931.png)

>>25592 Im insanely close to at least figuring out XL's quirks but i can't even find a good stocking anarchy lora for XL that doesn't force the gainax style. Currently crippled by that. The one on civitai is impossible to tardwrangle. this is sorta why i let myself be so comfy with 1.5 until now, it was such a gitterdone model that wasn't picky with prompts and settings.
>>25593 yeah, XL responds to prompts much better than 1.x ever did. and the higher resolution helps a lot with lora training, if only pony didn't absolutely destroy their TE
>>25594 I dont know about the prompt adherence, 99% of realistic models ive tried were way more restrictive about detailed poses in prompts and stuff, though it may just be samplers or some cfg fuckery i dunno 1.5 was never that temperamental, just limited in its capability for complexity. Like in the alice pic i just shared, no WAY could i ever achieve that with my OG 1.5 setup, it couldn't handle anything more than a simple shot with either a character and no background, or just a background. XL is already soaring with that despite issues.
>>25595 ah, we are talking about entirely different things here then. I strictly gen anime, so I only ever use booru prompts
(9.59 MB 6144x1164 xyz_grid-0000-552288701.png)

(1.32 MB 1916x305 dfgbzfdsgb.png)

i genuinely do not understand SDXL
>>25593 >>25587 Btw anon you seem to have jaggy lines that are caused by default sdxl vae Some of the models have the sdxl0.9 vae(good one) baked in, some don't to avoid the headache I suggest you using the good vae all the time https://huggingface.co/autismanon/sdxl_vae/resolve/main/sdxl.vae.safetensors?download=true
>>25598 thanks for telling me, now i can redo this other gen test i was working on with that vae.
are links okay here? gonna post a google drive.
>>25600 yeah, I post my personal mega drive of lora all the time, it's pretty much fine so long as it doesn't link to anything against the rules of the board
>>25601 Alrighty, just needed to make sure. Anyway, Here's the PC98/ElfPC lora i tried to make. This was the intended result >>25390 https://drive.google.com/drive/folders/1qqBFuV4fSuNen2ny6H2voy-_hSK0onW8?usp=sharing all 10 epochs, civitai says the 10th is the best, im starting to think i just didnt train it nearly enough on acutal pc98 and elf-pc corp stuff + my own gens might have fucked the overall quality given it's from 1.5.. ugh. If anyone wants to give this a shot and help me figure out how to get that result that'd be real fuckin neato. Might just try to grind another 500 buzz to give this another shot, add some more to the dataset..
>>25589 Some of those folks joins WD team?
>>25603 No, but the WD team has had a fairly open dialog with them for a while now.
What's the current status on local? is it still better to just sub to novelai for the time being?
>>25605 mixed bag. pony is manageable but NAI is better if ur not looking for niche
>>25606 from what I've heard, Pony is compromised regarding characters/artists due to drama?
>>25605 >for the time being? Until the Cascade finetune comes out, yeah. But then again we're probably closer to NAI V4 than the Cascade finetune's completion or whatever WD is doing. >>25607 >Pony is compromised regarding characters/artists due to drama? Aside from being poisoned at a fundamental level (mistagging, mixing -mistagged- sources, score gradients that somehow got mixed up, etc etc etc etc etc etc etc) that troon decided to axe 2hus and vtubers due to either spite (read: having personal beef with their respective communities) and he "dropped" artist tags aka he removed most of them (some slipped through the cracks and didn't get obfuscated) and the ones he absolutely wanted to keep using got hashed by some sort of algorithm. Anons also found unprecedented amounts of scats, apparently "ces" (which gets picked up as feces) is the strongest token in the entire model iirc. Oh and tons and tons of what anons think are pics of homemade fleshlights (that's "gpo", not sure what it stands for)
>>25608 ponyfags never disappoint
(2.41 MB 1280x1920 00081-4128618511.png)

(2.56 MB 1280x1920 00089-2412173109.png)

(2.43 MB 1280x1920 00100-605015453.png)

(2.50 MB 1280x1920 00087-1516353573.png)

>>25607 its "shit" by default but still the only decent thing, people use loras with it and a spin-off model to unfuck it mostly
in case you still browse the 4groids thread /^\d+-\d+\.(png|jpg)$/ /^\d+\.(png|jpg)$/ /^\ComfyUI_\d+\_\.(png|jpg)$/
>>25613 what's the d supposed to be filtering? that seems like it'd hit alot of false positives
>>25616 Forehead sex... I guess I find large foreheads erotic enough to insert
>>25616 and they said nai can't do gpo
>>25619 cute pantsu
>>25614 Comfy posts and forge posts that use the default naming scheme, groid switches between them now
>>25621 >>25622 >>25623 yay! >>25625 Cute brat! I have too much free time recently and I literally can't stop genning now.
>>25626 needs cute micro bikinis
>>25628 Perfection.
(1.45 MB 1024x1024 00238-4163447572.png)

(1.29 MB 1024x1024 00239-3586646631.png)

(2.29 MB 1224x1536 00241-2692475651.png)

(1.86 MB 1120x1408 00245-1845160460.png)

even when the model is made to fucking generate perfect images in 1-7 samples, i still cant get anything consistent or on par with 1.5 im 100% convinced there's a ghost in my machine that doesn't want me using SDXL. God DAMN. https://civitai.com/models/311157/vxp-xl-hyper
(972.21 KB 1024x1536 00156-1360294940.png)

(1.02 MB 1024x1536 00143-1559628964.png)

(1.23 MB 1536x1024 00264-52276265.png)

(1.03 MB 1536x1024 00198-1915665683.png)

this place is like the only place i can go to once a week to vent about XL and at least i dont have to see BBC spam/retarded gay shit/bot spam, somewhere that isnt reddit or twitter or a discord, and i really like that butyeah sorry thats like all i do really heres some good gens i got from 1.5 this morning.
>>25631 I do wonder why you have so much trouble though. I find that SDXL is a much easier model to work with in terms of genning, training it's a bit more of a bitch, but the results are worth the effort usually
(1.10 MB 1024x1024 00256-1124276.png)

(1.14 MB 1024x1024 00257-2447710423.png)

>>25632 i will literally copy gens verbatim from HF and 99% of the time they do not come out right here's one i just did a few mins ago. It's like 1.5 tier. https://civitai.com/images/13428661
>>25631 I haven't watched PSG in ages, time to rewatch it.
(1.20 MB 1536x1024 00265-554012951.png)

(721.99 KB 1152x768 00266-1039947100.png)

>>25633 >HF meant civitai >>25634 honestly i've been autistically obsessed with stocking since i was 14, but i'd never watch that piece of shit anime again kek it's pretty bad if you're not just watching it for the fun action scenes.
>>25633 hmm, odd, I dunno, I haven't really been finding any issues with SDXL prompting, things have been looking pretty distinctly better for me on SDXL based models, namely autismmix confetti
>>25635 had a hell of an OST too
(791.02 KB 1152x768 00267-1718289491.png)

(735.09 KB 1152x768 00268-2561800783.png)

>>25636 autismmix always seemed to work better than base pony, i think thats one of the merges that fixes the retarded score prompting. >>25637 oh god man thats the best thing to come out of psg besides stocking herself, Champion is the best anime track ever. by the way fellas what does it usually indicate when switching to a VAE causes this grayness issue? is it incompatibility? clip skip? the style's suddenly changed when i set to clearVAE and i kinda like it.
>>25638 >oh god man thats the best thing to come out of psg besides stocking herself, and kneesocks.
>>25638 yeah, it stabilizes pony a bit, it's all I use, because it works well with the lora I make, well mostly. I don't use the score or source tags personally
>>25632 I cant into autismmix either, idk. Most of it just ends up looking like shit
>>25642 I just use style lora, I have like 40 that i've made to choose from
>>25640 >>25642 >>25643 >people fucking prompt like this with SDXL models/Pony and get better results than i do what in the fuck. also ashamed that most results of stocking are all futa/futa on futa.
(615.98 KB 1865x886 dzcfb zdfbdzb.png)

(833.24 KB 1024x768 00319-2113136247.png)

>made sdxl so mad that it spat out a sketch all i did was remove that schizoid negative from this example >>25644 wtf
>>25608 I guess i'll stick to NAI till Cascade releases. a model without 2hus or chuubas is basically a nonstarter
(1.56 MB 1024x1536 00359-346857210.png)

(1.58 MB 1024x1536 00357-3746329475.png)

(1.61 MB 1024x1536 00348-1172415588.png)

(1.74 MB 1024x1536 00342-181563540.png)

AutismMixSDXL Lightning (with a metric fuckton of steps kek) wins. After all this fenagling, and niggerdallying, It came down to bad LORAs and promptlet issues. That's all. This image from Civitai just delivered me from evil. https://civitai.com/images/13323437
Anyone want to train the loli from wuthering waves? Absolutely nothing on civitai or anywhere and makes me sad
>>25650 i'm training hina from BA right now, but I am totally down for it
went through autism's civit gallery looking for prompts to crib and boy do i regret it
>>25652 >actually looking through gallery I already need eye bleach scrolling through LORA preview pics Going through gallery is willingly cursing yourself
>>25651 Nice Looking forward to it
>>25654 seems like with gel + pixiv diving I managed about 40 images, so it might actually turn out fairly good, will take some work though
(1.50 MB 1024x1024 00085-3582932726.png)

(1.60 MB 1024x1024 00086-2465757884.png)

(1.72 MB 1024x1024 00087-3616755130.png)

(430.16 KB 919x722 im gonna have a cornary.png)

>ghostxl knows the pc98 and dithering style natively >I had no idea all these months >trying to brute force a niggerrigged civitai lora >when i couldve just prompted this with these basic settings this whole time im. going through the stages of death right now, but i just blasted past bargaining and denial.
>>25656 sovl
(1.57 MB 1024x1024 00124-1672906489.png)

(1.32 MB 1024x1024 00174-1152086643.png)

(2.04 MB 1280x1280 00180-1498206082.png)

>>25657 been at it all morning now, i have to restart my pc or forge every so often because my gen times start to tank, but i'm really starting to get a hold of how this works. ..but ghostmix is not a great model for XL standards, anatomy accuracy is all over the place. At the very least its very consistent for results. >and so far not a SINGLE other xl or pony model i've tested actually knows the style at all, doesn't recognize pc98 or dithering whatsoever
>people still clinging to the absolute shithole that is 4ch*n instead of posting here
(1.35 MB 1760x484 catbox_5odld2.png)

(3.16 MB 1664x2288 catbox_zw83wv.png)

hina seems like it turned out pretty good, I'll resize and upload it soon
(1.35 MB 1024x1024 00216-3808237135.png)

(1.51 MB 1024x1024 00217-1516513426.png)

(1.36 MB 1024x1024 00218-881827104.png)

(1.61 MB 1024x1024 00221-2953320877.png)

fucking FINALLY after months of toiling and horseshit figuring out how XL works, i nearly 1:1 recreated that nice 1.5 elf-pc pc98 style. ..but these xl models are pretty, well.. fucking garbage lmao c'mon, this is barely any better than 1.5. It's just not horrendously awful at bodily consistency. The hands are still bad and the eyes are grainy. Anyone have some good hand loras or something i can do to alleviate this? I really can't manipulate my prompt any more to improve it. Also clones, that's random.
>>25662 aand site ate the last image
>>25662 >>25663 You are using pony tag structure on an SDXL derived model. Turbo is also known to be well, more shit, quality wise. I don't want if to go ahead and explain more at risk of sounding patronizing. Are you fresh off the 1.5 boat?
>>25664 honestly my fucks to give ran out so fast i didn't even notice i still kept those gay ass pony tags in my prompt they didn't affect the prompt gen for sure though, the model ignores em. nah you wont sound patronizing for just trying to help, ive been asking for help in this particular basketweaving forum for a while, off and on anyway. That Turbo was unironically better for what i was looking for than the myriad of models ive tried so far, at least 6 different pony mixes and pony base as well.
>>25665 I see. I just came here from /h/ after someone dropped the tip there's an hdg board here. Place turned into too much of a shithole to be tolerable. >That Turbo was unironically better for what i was looking for than the myriad of models ive tried so far, at least 6 different pony mixes and pony base as well. Are you looking for lite 2.5D or what in particular? Pony has a strong base of many styles all garbled up, it's probably who a few adequate loras can reshape your gens style rather efficiently. As for anatomy stability, I'd recommend amix DPO, it's a good middle point between composition flexibility and anatomy quality. But personally I keep several models at hand and swap them depending on what I'm trying to gen. I can give you some tips for quality enhancement if you want. But my experience is mostly with amix/pony
>>25666 nta but welcome.
>>25666 I'm trying to do this here >>25587 >>25656 >>25425 >>25390 then i have a few posts were i experimented >>25392 >>25398 among others Honestly in retrospect i really shouldve trained that LORA for PONY instead of base XL.. fuck. I dont have the energy/tism to grind for more buzz on civitai to do it.
(2.22 MB 2816x2560 00163-1756041244.jpg)

>>25667 I appreciate the hospitality, m8 >>25668 Ah you're a baker. Alright, that's a bit above my paygrade. I thought you were a late blooming 1.5 baby >Honestly in retrospect i really shouldve trained that LORA for PONY instead of base XL.. fuck. I dont have the energy/tism to grind for more buzz on civitai to do it. Probably, it's the default base model and from what I hear Lora training is easier, but I dunno, I am purposefully not diving into that rabbit hole because genning is too time consuming as it is. Gen unrelated
(727.41 KB 1024x1024 00063-2645270261.png)

>>25669 what would a baker be? I would absolutely consider myself someone still sorta new to XL, i focused mostly on 1.5 for a while for speed reasons, but recently XL's become accessible enough. I'm doing everything on a GTX 1080. I do believe in the idea of simpler prompts for better results, but.. ive been sorta growing nihilistic after seeing all these god tier gens on civitai with horrendously braindead prompts. I think some models are just really generous about it as long as the settings are fine. >this sort of gen for example is what kept me sticking to 1.5 for as long as i did
>>25666 Glad the tip was enough. Shame what happened to that place. Wasn't always like that back in 2022.
(1.96 MB 2640x2240 00012-1644712688.jpg)

(2.08 MB 2304x2816 00563-2441932596.jpg)

(2.46 MB 2816x2304 00025-2540012612_nt.jpg)

>>25671 It's been on a downwards spiral. It gives a bit of credence to those theories stating it's deliberate to lower the overall site quality and make it unusable. Janitors banning me for "off-topic" (lol) while leaving the rampant botspam untouched really rustled my jimmies for good. >>25670 >what would a baker be? A lora trainer >i focused mostly on 1.5 for a while for speed reasons, but recently XL's become accessible enough. I'm doing everything on a GTX 1080. 8gb? tough but not unmanageable, especially with tiled vae and such. You're on Forge, correct? >I do believe in the idea of simpler prompts for better results, but.. ive been sorta growing nihilistic after seeing all these god tier gens on civitai with horrendously braindead prompts. I think some models are just really generous about it as long as the settings are fine. Well, it's true that pony handles 1girling and simple concept prompts rather effortlessly. But then again you never know how much gacha went into a particular gen. 1girling is so easy now it's become superfluous for me. Switching to pony raised the bar quite a bit so now I almost exclusively try to generate full scenes with contextual background and several characters, or just plain weird shit you don't see very often. Maybe I just like to gacha, but I know 1.5 would fuck me sideways when trying to gen obscure stuff or getting (almost) every detail right simultaneously. Pic related, some older gens
>>25672 yeah forge. Only UI i could ever get working right besides Comfy, which is mostly just annoying to use lol. Honestly if i could somehow train LORAs on 8gb of vram id be trying, but i hear it takes literally days. 1girls are too easy, way too easy. I tried some more slightly complex stuff and it was tough, that was one of the main reasons i knew i couldnt cope with 1.5 anymore.
>>25673 yeah forge. Only UI i could ever get working right besides Comfy, which is mostly just annoying to use lol. I assumed. Forge is well optimized off the bat. It can get a bit lighter with some settings, but at slight inconvenience. Frankly other than efficient resource management I think the reason Forge takes the cake is how it handles Loras. A1111 is abysmal when model and lora swapping. >I tried some more slightly complex stuff and it was tough, that was one of the main reasons i knew i couldnt cope with 1.5 anymore. Yep as I said, raises the bar. Can't go back. Hell, I even feel limited with the shit pony + concept loras can do because I'm so spoilt. I mean at some point you can just go full controlnet and so whatever you want, but it sort of takes the fun out of it. Anytest is pretty fun to dick around though, it's like controlnet for dummies, very fast and simple.
>>25674 So, with these LORAs, what are my XL option? Or are there any specific PONY models that can still load the LORAs and shit just fine? https://catbox.moe/c/b6l4lp those hands are pretty 1.5 tier. among other problems with some older tests ive included in this album.
>>25675 How do you mean? what are those loras made for? As far as I understand it the lora must be trained on the base model to work properly. Basically if it's made for 1.5, I THINK you can still load it on Pony (at least in some cases) and it works (badly), but it takes like 20 minutes to process. Seen some people do it, haven't tried it myself. If they're SDXL based I think they'll work but suboptimally (haven't tried either, I stick to pony Loras). Basically you should always use pony loras for pony derivate models (pony, amix pony/confetti/dpo reweik, naipony, etc) and sdxl loras for sdxl derivative models. I don't know how far this rule bends though. Civit has a neat filter option to include and exclude models. Although civit loras tend to suboptimal (not sure why I'm a stranger to baking, just know very often they mix like shit and don't wanna play ball). I'd check loras under the same artist/theme. Or were you asking for recommendations of similar looking stuff?
>>25676 the character loras and that style, the whole point is to get those characters and that style. Like the posts i linked >>25668 Ive just been trying to get that look in Pony/XL, so far the problem is as youve pointed out the pretty bad fingers/anatomy. Id try pony but none of my setup is compatible with it.
>>25668 >I dont have the energy/tism to grind for more buzz on civitai to do it. NTA, but if I got it correctly you just trying to make pc98 style woking correctly on pony? Just drop your dataset here, I will give it a shot tomorrow
(2.59 MB 2816x2560 00211-2295023274.jpg)

>>25677 Yeah, think you just gotta do a painful full model switch man. >>25678 What a lad This place is comfy. I think I'll be posting my overengineered trash here from now on. Like this one
>>25678 Damn id be forever grateful if you gave it a shot the main folder is all the original images, some new ones, the subfolder is the dataset from what i sent to civitai, tagged. https://drive.google.com/drive/folders/1MxQeQE7iaYRUtZWbfj14Sf0yhXOeBdzD?usp=sharing >>25679 we'll see if that madlad can achieve what ive seen as impossible. that was a fun time of experimenting and shit though.
>>25680 Not surprisingly that you got very bad results using this dataset, it's first of all inconsistent, just look on the first two and last two, second is you need to upscale to 1024+ res for xl models and third it's just too low, only 17 valid images I could find, if you can grab more from anywhere, would be a lot easier, trying atm to push those 17 pics into the oven but I honestly dont have alot of hope for a great outputs at the end
>>25681 would it make more sense to only train on the shit i made in SD itself? i figured thatd be even worse because of all the bad anatomy inputs >totally forgot that awful hand-foot one you just posted oops
(1.89 MB 1248x1824 catbox_u3ufjr.png)

(1.51 MB 1248x1824 catbox_sxixml.png)

>>25682 No, it's better to upscale the rest of the images from game/artist/whereever-you-take-it-from, exactly what I did. You should avoid putting 1.5 gens in your datasets for as long as possible, unless you want another aom2 model, with a good gens made with pony, which don't have problems like the limbs at your example and represents the style for iterative learning it's not that bad idea to start adding them in a dataset for a second run with bigger dataset Here what I've got, with some catbox examples https://files.catbox.moe/duqx3r.safetensors
>>25683 huh i see. Alright then. In that case, I downloaded this 1,000+ games CG set from nyaa.si, literally thousands of cg's from all sorts of games, for now i'm gonna narrow it down to just trying to recreate that elf-pc corp lora for 1.5 so it's easier. Updating that folder with the elf set. >its 2 thousand cgs kek
>>25684 Just 200 best would be enough, couldn't do more atm, have to leave
>>25685 its all good, tomorrow's another day. I need to sift through the dataset anyway and figure out how upscaling works, I didn't touch it for the LORA initially because from what i understand, it gets auto upscaled.
>>25686 its even worse idea to let sd-scripts make auto-upscale, the best you could do is just use some dat upscaler in forge or a1111
>>25687 if he's doing it specifically because he wants dithered art, wouldn't plain integer upscaling be the way to go
>>25687 understood. first thing in the morning is some scaling on my own >>25688 i would think, yeah. i want that crispyness.
>>25688 How about adding loss mask border to low res gens? It will change composition of course as AI will have to think of something to outpaint the masked out area, but you will preserve all the high frequency detail of your dataset this way and won't taint it with upscalers shitty high-frequency detail. This will probably work terribly though if most of the dataset is low res.
by the way one question i need to ask before i hit the hay; when trying to make a style lora, the actual exact contents of the images wont matter will they? the big concern i have with anything 90s anime related, in this case with pc98 especially, is pussy censorship. I mean in some of the best examples of this dataset, there's flat out no pussy. Yet that elf-pc LORA i was using from civitai has no problem making pussies, thought that was interesting. I figure pony shouldnt have a problem but i thought id ask.
>>25690 To low res images* gens=images in my mind now apparently >>25691 Just tag them properly. Vagina isn't drawn - don't add the pussy tag. Censorship? Add censored, bar censor, mosaic censorship, whichever applies. Having at least some properly drawn vaginas in the dataset would be better of course.
(2.88 MB 1664x2288 catbox_6jnkw0.png)

(3.07 MB 1664x2288 catbox_76bj49.png)

(3.24 MB 1664x2288 catbox_clli5c.png)

(2.70 MB 1664x2288 catbox_7us64a.png)

>>25650 v1 has been baked and tested, it's... not 100% consistent, the outfit is like 80% of the way there most of the time, but it does sometimes get it right 100%. it's fairly generalized though, so you can fairly easily put her into other outfits, her actual character traits are accurate though, so I'm releasing v1 for now, and I'll work on it some more later https://mega.nz/folder/OMY2RRxQ#zYaokRZV361WSNImfrBCrQ/folder/6chDzRaZ
>>25692 alrighty
>>25693 Nice, I'll try it some now Thanks a lot
So on the subject of scaling low res images to make an XL style lora, when i'm trying to scale, Am i meant to make them exactly 1024/1024 or at least the closest they can be? Width wise is easy but height will always be a weird resolution given how that system worked.
>>25695 No problem! I hope it's up to standards, the lack of good art made it somewhat hard to make in comparison to like hina for example (Which had 905 images in my dataset)
>>25693 cute
Is there any real way to speed up upscaling besides the model? Been repping 4x ultrasharp for a while because its the fastest model i know while also being quality, but every damn upscale is just so resource intensive it seems.
(1.21 MB 1176x1176 00036-491883404.png)

(1.22 MB 1176x1176 00037-91867379.png)

(1.19 MB 1176x1176 00038-557921966.png)

(1.21 MB 1176x1176 00039-2671722489.png)

heres some "style accurate" stockings as payment for all my sperging and needing help lately. you guys are real ones.
(1.16 MB 1024x1024 00078-3522906622.png)

(1.14 MB 1024x1024 00068-3522906620.png)

(1.16 MB 1024x1024 00065-3522906617.png)

(1.14 MB 1024x1024 00071-3522906615.png)

>>25701 so so good, fuck.

(1.13 MB 1024x1024 00098-211438758.png)

(1.12 MB 1024x1024 00097-211438757.png)

If only i had toughed out the rough first month of trying to use SDXL instead of giving up, I could've made gemmies much earlier. Holy fuck. Pony's great. I kneel. >I will then ascend beyond mere 1girl limitations soon enough This is base autismMix pony by the way.
(1.14 MB 1024x1024 00174-1247660962.png)

(1.11 MB 1024x1024 00173-1247660962.png)

(1.20 MB 1024x1024 00052-3739997144.png)

(1.14 MB 1024x1024 00062-3522906618.png)

Last ones for today/forseeable future, I don't wanna be annoying. I find it interesting how easy it is to get style variations, Though i'd really love to try and stick to one, I don't know how to tardwrangle pony to do it besides seed variation trickery.
(2.14 MB 2816x2048 00148-2598856977.jpg)

>>25704 Most straight forward way is lora mixing, although controlling style on pony has always been a core issue even then. Data for poses is all over the place. So say I'm using a model train on ecchi stuff and I try to use it with hardcore, then the style becomes weaker. Personally I try to combine several flexible loras of similar style at low weight to mitigate this, as to feed it broader data
>>25705 "For now"
>>25705 you've been genning for so long im surprised you can still feel anything when you see her
>>25706 Ahhhh that explains why im starting to see some gens on civitai specifically under Pony that combine some very specific styles i figured made no fucking sense. Now this makes sense https://civitai.com/images/11642164
(2.30 MB 2304x3072 00265-3406103130.jpg)

>>25709 Yeah. Lora mixing is sort of an art itself if you want reliable mixes. Plus a lot of loras themselves are trash to begin with. There's no way to really tell for sure but through trial and error, soon enough you'll single out some loras as more reliable than others. This baker on /h/ has some pretty good ones. I check on his mega every now and then. On civit there's an incredible volume of stuff, but it's completely unreliable quality wise, a lot of trash.
>>25644 >>25669 >>25679 >>25706 >>25710 catbox????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????
>>25696 More than 1024x1024 size, if h side is 832 w side should be 1216 or bigger for example, etc >>25680 So looks like you still haven't gathered new dataset?
>>25719 oh yeah check gdrive, i picked out a particular pc98 dev, then went through and picked 200 images out of that. now i gotta resize the images then do the tagging.
(823.00 B 177x21 dzgbvdazgbv.png)

>>25719 literally had to figure out forge even has these tools, uh.. is this too huge of a scale? i hit it to 5x scale and this seems to be the biggest size.
>>25719 dataset subfolder's been created in the link i shared, images are all resized and tags have been provided. >i hope i did it right
>>25722 It looks fine, I will try to bake with that, the only concern is there absolutely no pussies I could find, its heavily censored >>25721 As long as it surpass 1024x1024 pixels it should be fine
>>25723 for pony based lora, censorship isn't a huge issue so long as you have it properly tagged and use "uncensored" in the positive prompt
>>25723 yeah thats why i asked earlier about no_pussy. i made sure to remove the vagina/pussy tags as instructed. >>25724 good, thats a relief.
>>25724 Yeah, I know, it's just great idea to show ai how to draw something, not just let model generalize it from basic knowledge I mean
>>25726 oh no, I agree, just noting that it should generally still work
>>25726 At the end of the day though, Pony is hailed as the god of all stable diffusion models and should be able to draw a pussy from its own memory without being taught. I mean, ghostmix from 1.5 was able to do it just fine. That was 1.5, in fuckin probably less than 512x resolution, and my shit was upscaled 5x. kek.
>>25728 Well pussy is not a problem it seems, but it really tends to almost half-cutting the top of the head from the image very often, because of dataset that have enough of that too, try it yourself https://litter.catbox.moe/5sihql.7z
>>25729 aw fuck i figured that might be an issue but i wasn't sure how pony would handle it. By the way, what settings are you using to get your results? Mind catboxing the images too? now, would.. replacing a bulk of the images with the top head crop help, or, just adding more images that don't have that kind of angle on top of the existing data fix it?
>>25730 Images there in the archive. Ideally it is better to have just full pictures without cropping near the head, replacing most of them should do the trick probably too
(1.14 MB 832x1216 00000-1382714565.png)

(824.26 KB 768x1024 00001-1382714565.png)

(150.78 KB 1870x391 awgawzgwa.png)

(345.52 KB 1843x879 sigmaballs.png)

>>25731 Im gonna go through the folders in a few mins and start picking out images, either swapping or just adding to the dataset. So far i think what this needs too is more different body types, maybe. Maybe more lolis. Gonna see. I really appreciate your work so far, I'dve had to wait months or a year to get better hardware before i could do this myself. interesting to note that changing the resolution to something more close to XL's standards changes the crop issue in weird ways, trying to recreate the first stocking image with the og seed also has the crop issue but worse I think i might do a complete re-install of SD forge, it's really weird how many times i've tried to remake someone's seed on my end and it never turns out right or even worse.
>>25732 that could depend on the RNG type, if it's GPU you won't ever be able to replicate it unless you have the exact same GPU they have
>>25732 >>25733 Not just this, there is also vectorscope enabled, so you won't be able to replicate it without extension
>>25734 ok, I have no clue what vectorscope is, shows how much I gen. i'm good at training at least (and I enjoy it)
>>25733 Might be, though i have other issues besides the RNG, changed it from NV back to GPU and back again, seemed to plague both. >>25731 By the way totally forgot to mention, for the activation text, i wanted to use a totally original text string instead of 1990s style or pc98 or anything, that way the base model can't influence the style, and i can apply the style LORA the way i'd like to without changing things like the way hair or tits are drawn. The images are labeled "P98E" since that's what i intended the activation text to be, should that be added into the tags, or just added to the model in webui? Im still pretty confused by how that works kek. >>25734 what's vectorscope?
(243.77 KB 919x839 crop P98E.png)

Answered my own question, it's just the feature in webui. "P98E" and "dithering" work perfect. LMAO thats such a horrendous crop problem, my bad man. Im so sorry i just wasted your compute on a slightly poisoned dataset. fuck.
>>25736 >By the way totally forgot to mention, for the activation text, i wanted to use a totally original text string instead of 1990s style or pc98 or anything, that way the base model can't influence the style, and i can apply the style LORA the way i'd like to without changing things like the way hair or tits are drawn. The images are labeled "P98E" since that's what i intended the activation text to be, should that be added into the tags, or just added to the model in webui? Im still pretty confused by how that works kek. Probably don't even need any prompts, but I just used every tag you added with autotagger there in the examples >>25736 >what's vectorscope? >>25735 You just can't fix everything in pony with training, it's a handy thing, don't need any shitty noise offsets for training with it https://github.com/Haoming02/sd-webui-vectorscope-cc fixing the dynamic range in the inference time
>>25738 oh that's cool, yeah, I just don't gen much, so I didn't know about it. I actually use the noise offset and pyra noise because it seems to help the model learn a bit more, in my experience at least. I don't know how you train, so I can't really speak for another, but it works well for me at least
>>25739 All I could recall about classic offset is it a horrible thing in perspective of using it with something else trained with it too. I just trying to avoid any offsets, besides maybe pyramid one, for training, and found that inference time noise scheduling changing can also boost style coherency with that extension, if you manage to tweak the params close to how artist draws colors
>>25740 that's pretty cool, Since I don't really gen, I don't use other people's lora, so i've been training with use along my own lora only pretty much. that being said, the noise offset I use is 0.0357 which is the same noise offset for SDXL, so it shouldn't even affect anything in this case
>>25707 You are correct... >>25708 >you've been genning for so long That is true. It is a good thing that I have 3 different chocos to choose from. Well I have a hard time getting bored of this brat and if I do, off to textgen I go.
>>25742 very sob worthy as usual
>>25738 Files are up in a new subfolder, i removed a good chunk of the dataset after realizing it was probably a boneheaded move to heavily train on ONE game, and that ONE game had the worst cropping of every game in the dataset, so that got trimmed down to a near handful-minimum, added different shots from other games. Hopefully this should do it, seems like however you train is foolproof, it's just down to a quality dataset. Quality over quantity this time.
>>25744 >Hopefully this should do it, seems like however you train is foolproof, it's just down to a quality dataset. Quality over quantity this time. It was always like this, hyperparams means nothing when you got shit input, you can't create something out of nothing, and when you have perfect dataset, hyperparams started means almost nothing, but yeah, it's really bulletproof when you are min snr with dropout and locon. I'll train with updated dataset tomorrow >>25693 You probably want to check this out https://github.com/KohakuBlueleaf/LyCORIS/issues/167#issuecomment-2132501197 Additional parameter would be handy now in easy scripts for old behaviour, when kohaku almost fixed the speed issues
>>25745 ah cool, i'll make sure to update it later, i'm currently grinding through levels on wuthering waves
(881.86 KB 832x1216 00014-1382714565.png)

(1.23 MB 832x1216 00016-1382714565.png)

(1.06 MB 832x1216 00028-1382714565.png)

(1.07 MB 1024x1032 00029-1382714565.png)

>>25745 cool man thanks. Honestly having a blast just playing around with the one you already did, gettin really fun results. Hopefully the better dataset will completely fix the issue.
(1013.75 KB 1024x1024 00071-1062827372.png)

(903.31 KB 1024x1024 00068-1062827369.png)

(967.46 KB 1024x1024 00067-1062827368.png)

(830.25 KB 1024x1024 00066-4083443311.png)

posting some alices made from anon's latest lora
(1.06 MB 1024x1024 00084-804971222.png)

(727.58 KB 1024x1024 00083-804971221.png)

(774.75 KB 1024x1024 00082-804971220.png)

(1005.04 KB 1024x1024 00081-804971219.png)

(1.77 MB 1344x1728 00004-604551891.png)

Heavy-bottomed and proud.
(1.04 MB 1024x1024 00164-2843668088.png)

(878.77 KB 1024x1024 00155-3957397260.png)

(1000.91 KB 1024x1024 00129-322365991.png)

(915.21 KB 1024x1024 00122-627702159.png)

holy fuck im having a ball
(1.10 MB 1024x1024 00174-3254137869.png)

(1.09 MB 1024x1024 00175-621622561.png)

(1.06 MB 1024x1024 00176-621622562.png)

(1.12 MB 1024x1024 00177-621622563.png)

my heart just exploded at how adorable this turned out
>>25752 distinctly not a fan, to be honest
>>25752 bruh
(2.43 MB 3072x2304 00012-1157639919.jpg)

(2.53 MB 2816x2560 00021-3620010912.jpg)

(2.57 MB 2816x2560 00123-1500103319.jpg)

(2.52 MB 2816x2560 00004-1500103323.jpg)

One of the side effects of pony using a furry base is monsters, humanoid and non humanoid are reasonably easy to do without lora aid. Fun stuff
>>25756 damn, nice stuff. catbox please
(2.53 MB 2816x2560 00003-401582663.jpg)

>>25757 They're more or less the exact same prompt. Bit gacha but not horrible Here you go https://files.catbox.moe/hq1489.png
(2.69 MB 2816x2304 00024-2968541912.jpg)

And a (zac) for good measure The more gacha your gen is seems to be directly proportional to inpaint gacha. Rather annoying
>>25758 >>25759 i like these a lot. there's something about cute girl pleasing freaky monster and they both look like they're having fun that does it for me.
(1.85 MB 1280x1920 00181-3420196013.png)

(1.95 MB 1280x1920 00170-410178188.png)

>>25701 hot, I should gen some more of that
>>25761 You should, but was I was unable to get the guy to cum in her panties.
(2.28 MB 1280x1920 00186-627587191.png)

(1.98 MB 1280x1920 00191-3872188598.png)

(1.81 MB 1280x1920 00209-1389091983.png)

(2.04 MB 1280x1920 00198-2447133826.png)

>>25762 I'll gen some later, this https://civitai.com/models/424332 helps make the concept work. But first, a few more 1girl + shiny holographic swimsuits
(2.42 MB 2816x2560 00332-3085926162.jpg)

(2.22 MB 2560x2816 00327-1675851958.jpg)

(2.16 MB 2560x2816 00329-3760920524.jpg)

(2.33 MB 2560x2816 00335-535076640.jpg)

>>25760 How about a monster pleasing the girl?
>>25764 nice! large insertion is such a based fetish
(2.17 MB 1280x1920 00340-1416020969.png)

(2.07 MB 1280x1920 00356-3870902958.png)

>>25763 The lora is definitely not perfect though. I wonder what kind of dataset they were working with
(289.60 KB 1022x439 etwatgawegaewgeag.png)

>>25767 holy fucking soul, thanks man. I really appreciate your hard work here. ..have to reboot my pc and everything because forge is giving me shit, but i can already tell this v3 is very promising.
>>25766 Yeah and it forces a style for me.
>>25769 All concept loras force style to different degrees unfortunately. There's no way around it because they cannot be possibly trained on everyone's style. A good concept lora will be trained in a variety of broad styles though, so you won't feel the style drift as much. I recommend using concept loras only if you really have to and with as little weight as you can get away with. Concept loras are mostly OK for inpaint though.
(2.75 MB 2560x2816 00366-2345347333.jpg)

Who else trashy appreciator
(3.59 MB 1664x2288 catbox_ffgldk.png)

(3.00 MB 1664x2288 catbox_us3hxv.png)

(3.03 MB 1664x2288 catbox_ni0igd.png)

(2.87 MB 1664x2464 catbox_nij7u5.png)

since I made encore, I decided to also try my hand at making verina, while this bake seemed to turn out fairly well, it has an oddity that isn't even in the dataset in which her head is slightly squished looking, so i'm gonna have to figure that out, for now, have some lewd verina
>>25772 Verina SEX
(1.81 MB 1344x1728 00026-15266613.png)

Could you tell, is this local or NAI?
>>25774 honestly Pony is so good, the only way youd be able to tell the difference is because of a gen'ers skill issue.
>>25774 Probably pony, since penis on cheek bulge
>>25774 local. custom res and nai struggles with some specific stuff.
(2.18 MB 1344x1728 00036-1584337015.png)

>>25775 >honestly Pony is so good I mean, the model is workable, mixes improves it, but omfg this model for LoRA training is the biggest piece of dogshit I've ever seen since I started on this AI image gen thing, it's so fucking painful to deal with as CBT with a chainsaw. Even at present days I can still see constant bakers complaining about how changing a brazillion settings, optimizer, scheduler, adjusting LR, trying new tech like masked loss, DoRA, LoRA+, and none of that helps to solve their issues with their bakes (Me included). They try so hard but still can't get the perfect LoRA, it always comes out flawed one way or another. I'm pretty convinced this model is completely unfuckable in that aspect. I still remember the good ole 1.5 days, at least the results LoRA baking were somehow predictable if you had some experience with it already. >>25776 Interesting way to spot it. >>25777 I myself used NAI and imo it looks kinda similar compared to these gens, didn't knew NAI has those issues.
>>25778 Well god damn buddy why don't you ask this guy >>25767 for a mentorship then? he seems to have pony LORA baking down to a 100% pure science.
>>25778 >Interesting way to spot it. There's actually other signs, such as smooth shading, filename and resolution. All of that could've been arranged as a trick to throw people off though. Cheek press is somewhat of an obscure thing NAI probably wouldn't handle too well though. I don't remember testing this particular thing it but NAI is overall a rather conservative model when it comes to concepts. Plus, if you were pulling an it's actually NAI it probably wouldn't be a flat style, it'd be something with complex shading, showcasing it without flatface (which is something NAI struggles with). But then again, without loras or access to the NAI model how would you fix that shortcoming? For that last reason, it's gonna be pony 99% of the time someone asks
>>25779 No man, this guy >>25778 is right, pony is absolute nightmare and fucked model, it was incompatible with any xl thing when it was launched, because of the raped with x100 lr unet and still struggles to precisely catch and transfer styles even between something close to it, like autismmix, but it's still the best we got atm for porn and uoh
>>25783 whats the style for the first two pics and that last one in the previous post?
(295.64 KB 790x1276 1423612138263-1.jpg)

(576.29 KB 967x1334 1571271337328.jpg)

>>25784 metadata's right there dude >"prompt": "loli, skinny, pointy breasts, oral sex, large penis, kissing penis, fellatio, cum jet, {{{{cum on face}}}}, one eye closed, cum on body, pinching cheek, male pov, amazing quality, very aesthetic, nsfw, [[artist:takayama kisai]], blush, wink, smile, artist:rustle, looking at viewer", "uc": "looking away, elf, intricate, logo wedding ring, backpack, tattoo, {worst quality, bad quality, {censored}}, amputee, deformed, fewer digits, bad perspective, bad anatomy, hair intakes, no pupils, @_@, interracial, condom, simple background, sketch, unfinished" plus a vibe transfer, probably one of these
>>25785 >metadata's right there dude no it aint, at least however you're getting it i don't have. thanks.
>>25786 most pictures in the thread have metadata other than the guy who just showed up and started posting jpgs
>>25787 oh shit whoops. im still new to this place, havent used 8 since..a very long time ago.
(2.89 MB 1664x2464 catbox_0e3fwg.png)

(2.78 MB 1664x2464 catbox_3cotdn.png)

(3.08 MB 1664x2464 catbox_ebm26q.png)

(3.32 MB 1664x2464 catbox_3j3av5.png)

>>25773 verina sex.
>>25788 no worries, i was in the same boat like six months ago
>>25770 One trick I would use in the 1.5 days: - Gen with the concept lora - Put the result in controlnet dept - Heavily reduce the weight of the concept lora But I haven't gotten good results with XL controlnet yet, could be a skill issue
>>25791 I can see that work better or worse depending on the concept lora itself. Even though you're using a guide, the base model still has to have (some) data within it to figure out what what should be drawn where. Here you are reinforcing it by >Heavily reduce the weight of the concept lora which isn't a bad idea. Moreover controlnet depth may still transfer some style/shading from the concept lora, since it's trying to trace back both proportion and detail. For example, I don't see a realistic 3D lora transferring optimal results into a lineart style. Because of this, and due to the fact it's a breeze to use you should give Anytest a try https://huggingface.co/2vXpSwA7/iroiro-lora/resolve/main/test_controlnet/CN-anytest_v1_fp16.safetensors?download=true As I understand it, it blends several controlnets together to get the overall composition, prioritizing pose and reinforcing things like canny and depth after (I believe). There's virtually no documentation sadly so no one's exactly sure how it works, just that it does. The biggest advantage is that it uses no pre-processor and is relatively light and fast, so you should probably keep it as a quick lazy option either way. Think of it as a composition transfer with flexible style and characters. If you're gonna try: select anytest as a controlnet model, upload a pic and you're set. It should be used with "my prompt is more important, weight range of 0.4-1.0 , depending on if you want an influence or to outright copy the source.
anyone got guides, civitai or not, on doing more than just 1girl? been grinding at Pony since like sunday and now im ready to venture into unknown territory.
>>25793 step 1: generate 2 girls step 2: inpaint them into whoever you want them to be, 1 at a time step 3: there is no step 3, it was a psyop
>>25793 I don't think you'll find any as it's an iffy endeavor to begin with. I can probably help you since I often prompt multiple girls and multiple people gens. There are 2 main ways to go about it: Gacha - Through unspecific prompting, more natural composition, less prone to errors but less control over your characters Regional prompting - Either Forge Couple or Regional, controlled characters but more prone to errors if the more characters you wanna cram into the same scene, plus extra trouble if the character's are invading each other's space (hugging, pressing against one another). Forge couple is technically regional prompting but way more straight forward and simpler to use, I'd recommend you give it a try https://github.com/Haoming02/sd-forge-couple For the first method, no addons required, you want to be ambiguous and let your prompt fall as it may. For example: >>25679 is gacha multiple girls. You may feed several contradictory options into the prompt so you avoid clones, like: brown hair, black hair, blonde hair. short hair, long hair, tied hair, messy hair, wavy hair, straight hair, etc. Clothing is best left unspecific, with tags such as: medieval, fur trim, emo fashion, casual, jewelry, loose clothes, armor. Since the individual detail will be randomized from a large pool of things, avoiding repeating accessories. The same logic applies to everything else, such as eyes, expression, and whatever else. Sex, there are options such as cooperative fellatio, cooperative paizuri, group sex, and so on. For background characters, success depends a lot on loras, some styles are good for them, some aren't. You want tags such as: people, crowd, crowded, others can help such as: faceless male, public indecency, exhibitionism, surrounded, audience. If you're crafting a really complicated scene, you can keep things such as expressions non committal and then correct with inpaint afterwards. Faces aren't as hard to inpaint as some other stuff. The second method separates the scene and directly controls the characters shown on each region. For example: >>25706 The easiest way is Forge Couple (link above). You can check the examples provided in the guide, it's rather straight forward. Personally I like to use custom separators, which you can define as any word you like. And the Global effect set to "first line" This way, you prompt as such: Lora stuff, Style stuff, Background and scene stuff, things affecting both characters, 2girls, _separator_ define girl on the left _separator_ define girl on the right What forge couple does is split the scene in regions equal to the amount of separators you use, either vertically or horizontally. You can also used the advanced option to set coordinates for what you want where. But at that point I'd just switch to regional prompting. There's more to it, as always, but you should start experimenting with these 2 methods, depending on what you require.
(1.26 MB 832x1216 00000-1191587899.png)

(1.35 MB 1024x1024 00001-626685205.png)

>>25795 thanks bud. I had to reinstall forge because its held together with bubblegum and cheap dollar store scotch tape, Forge Couple works great with the right prompting. Just gotta figure out LORAs and LORA characters.
(2.76 MB 3328x2420 00352-709301321.jpg)

>>25796 Neat A few notes on the limitations of Forge Couple: Bleeding (unintended transfer of one character properties to another), will happen on a regional basis instead of a global one. What does this mean? were a character to be rendered in the middle of the regional cross-section (e.g. in the middle of the scene on a 2 character prompt), then it will blend into a mix of both characters, sometimes more naturally, sometimes its hairstyle, eyes, color, breasts and clothing will be hard split at the edge of the region. This happens more frequently when using 3, 4 or more girls. Forge Couple is also ignored on i2i multidiffusers (I don't know about Hires and Kohya). But if you're upscaling with multidiffusers just bear that in mind, you may need to inpaint to fix the unavoidable bleeding after processing. Forge Couple is mostly stable on lineups. For complex compositions your chances are better with regular prompting. The lineups don't need to be frontal standing, it just means the characters don't invade into each other's regions. Picture related, a 3girls forge couple gen
(1.16 MB 1024x1024 00015-312567756.png)

>>25797 yeah i have no clue what i'm doing with the LORA prompting kek either new line doesnt seem to be seperating the characters like it should be by default, or i dunno. I just did the first two pics copying from the github's example, with normal not advanced.
>>25798 Use the custom separator, the lines can get pretty confusing [x] enable Region Assignment: Basic Couple Separator: _sep_ Tile Direction: Horizontal Global Effect: First line Here you put everything but the girls description _sep_ girl on the left _sep_ girl on the right For example <lora: mylora> <lora: mylora2> score_9, score_8_up, source_anime, 2d, outdoors, park, park bench, sitting, day, nature, 2girls, _sep_ 2girls, school uniform, vest, blonde hair, long hair, sitting, brown eyes, large breasts, _sep_ 2girls, casual, loose clothes, short hair, black hair, green eyes, small breasts, Using line separation is fine but it can get confusing with a long prompt. Try a test prompt without loras or anything and with 2 very different girls just to make sure its working properly, ramp up complexity after.
>>25799 By the way, the reason you use 2girls instead of 1girl is likely to play with region bleeding. If the characters are right on the edge, the region knows there's a sliver of a character in the region. You should think of what will be drawn on each region, considering overlap. For example: For 2 girls: 2girls (global) _sep_ 2girls _sep_ 2girls For 3 girls: 3girls (global) _sep_ 2girls _sep_ 3girls _sep_ 2girls For 4 girls: 4girls (global) _sep_ 2girls _sep_ 3girls _sep_ 3girls _sep_ 2girls However many characters are adjacent to the target character. 2 on the edges, 3 in the middle.
(1.26 MB 1024x1024 00023-626685205.png)

(571.84 KB 1871x966 hhhhhhhhhhhhhhhhhhhhhhm.png)

>>25799 >>25800 Odd it really just doesn't seem to be following the prompt consistently whatsoever.
>>25801 This is because on advanced you need to specify the global region (first, in this case) as covering the whole image So it should be 0.00:1.00 0.00:1.00 1.0 (global) 0.00:0.50 0.00:1.00 1.0 (left girl) 0.50:1.00 0.00:1.00 1.0 (right girl) For simple stuff such as this stick to Basic, only use advanced if you need it.
(90.98 KB 1280x768 example.jpg)

>>25800 Here, I don't know if needed but I made a diagram that should clear up why the overlap should be addressed in the prompt.
(580.88 KB 1854x955 ehnbrsfhnbtdrfh.png)

(1.26 MB 1024x1024 00030-626685205.png)

>>25802 >>25802 even then, its still only partially following non-lora prompts. do i need to be fiddling with cfg scale or anything to get this working right?
>>25804 Switch to horizontal and try again.
>>25805 I have no clue why that worked, thank you man. Now i'm down to relearning prompting just to get this working right kek, say i wanted to change this prompt so one girl is a futa, what am i meant to do? I tried just adding "futanari" to the 2nd one girl, didnt work, tried 1girl, 1futa change to the whole prompt, didn't work, then finally tried 1girl, 1futa, futa with female, it just changed the prompt to one character that wasn't futa at all.
>>25806 >I have no clue why that worked, thank you man. Because vertical is for Top area <-> Bottom Area Horizontal is for Left area <-> Right area You're welcome >Now i'm down to relearning prompting just to get this working right kek, say i wanted to change this prompt so one girl is a futa, what am i meant to do? >I tried just adding "futanari" to the 2nd one girl, didnt work, tried 1girl, 1futa change to the whole prompt, didn't work, then finally tried 1girl, 1futa, futa with female, it just changed the prompt to one character that wasn't futa at all. I'm not sure but I can guess. Change Global line to: 1girl, 1futa 1girl, 1futa, penis, erection, testicles, futanari 2girls, pussy, Something like that maybe, I barely ever tinker with futa, sorry. Also I'd recommend to get the danbooru autocomplete plugin thing. It's closer to optimal tagging, even though it omits some e621 tags. And, you're using a pony derived model, high quality, best quality doesn't do much at all. Score tags should do that.
>>25807 >You're welcome coffee will never truly fix whatever's wrong with my CPU, oops. also that's not really my prompt i just got it from here https://github.com/Haoming02/sd-forge-couple?tab=readme-ov-file i might grab that plugin though, would help. By the way while we're on this subject, there's multiple different plugins for speedups at higher res, what *should* i be using? Hypertile seems to work but i'm concerned about the quality tradeoff there.
>>25808 >also that's not really my prompt i just got it from here Ah yeah, that makes sense, my bad >By the way while we're on this subject, there's multiple different plugins for speedups at higher res, what *should* i be using? Hypertile seems to work but i'm concerned about the quality tradeoff there. I'm sorry I can't help you there. I am unfamiliar with performance stuff. Download more Vram? The only thing I can recommend is you lean on mixture of diffusers as a post process to upscale while feeding back full detail. It's worked for me rather well with 12gb, although I don't know what you're working with.
>>25809 I never OOM at 8gb but it could be slightly faster i'm sure. Gtx 1080 clocking 1880mhz consistently. 1280x1024 for example seems to be taking 1min 23s usually. I dunno maybe i should just be patient for now until i get a new gpu kek
>>25810 That's rough, but then again it could be worse. I'd recommend getting a cheap(er) mining GPU with as much VRAM as possible. Although if you're gonna upgrade I'd aim for 24 just to future proof a bit. Sorry I can't help a lot with performance tweaks. A lot of them come at a smaller or larger quality cost so I haven't looked into them.
>>25811 Honestly coming from LLMs, there's zero real reason to buy mining GPUs like the teslas anymore. They're showing age, speed/VRAM pros aren't really adding up anymore. No wonder though, given, well their age. I got lucky that the 1080 is AT LEAST holding up for now, but a year from now i won't be running it for A.I stuff for sure. 24 gigs seems like the solid place to sit, which is why i'm heavily considering two 3090s, probably for separate systems even. For now, even at that speed, I still can't hold back getting addicted to this shit. Like, PHEW, Pony is GOOD.
i need LORAS. give me your LORAS. >>25810 for comparison, a 2080 ti does about 1.8it/s at 1280x1024 >>25812 workstation cards are a pain. anything older than pascal isn't supported by CUDA anymore, pascal has dogshit SD performance, and tesla costs a lot of money if you don't have a special sxm2 setup. LLMs aren't particularly bandwidth-sensitive once the model is loaded so i wonder how competitive a pile of 3060s on mining risers would be, cause that would be about 50% more vram per dollar compared to 3090s.
>>25813 volta, not tesla
>>25813 problem with dual cards is it varies on how hard your performance improvement tanks relative to what youre running, so im beginning to think just running 2 rigs, and making sure they're power regulated reasonably and that's golden we're still years off, i think, from fully integrating voice + text + SD in any way. Like livewaifus and game integration, so two 3090 stations running their max potential i think is smart. that and electricity is a nonissue for me.
went through my NAI save folder and ended up with 172 artists i like and don't have loras for yet pony is hard to train for, you say
(2.23 MB 2816x2304 00070-3188051834.jpg)

Why are dragon girls so hot?
So what is the level of accuracy that Pony operates at? What level of anatomical consistency should i *actually* be expecting? It honestly seems pretty all over the place, at times a solid 4 out of 10 gens has bad hands, maybe 1 in 10 gens gets extra limbs. Without any negatives, using only score_9. Could this just be a LORA issue?
>>25818 Yes it depends on loras, style tags, pose-composition, prompt load, everything mixed together. I mean even your sampler will have an effect on errors if you stress the hell out of the composition by cramming 200 tags in your prompt. So there's no answer other than it depends, since every setup is its own particular case. All that said, on average, it's currently more stable than anything around, although not magic. Use DPO for more stability.
>>25818 >>25819 Another thing, Loras (unfortunately) are mostly inconsistent garbage and often can't handle anything but the bare minimum. This is probably the baker's fault, although sometimes it's understandable if the artist itself lacks certain themed data. Whichever case, there's a lot of people out there just shitting out 1girl "good enough" loras. I've downloaded tested and deleted 80% of loras downloaded from civit. Worst case, some of these bad loras can completely destroy your gens if you use them with a nono tag. When in doubt lock the seed and test same gen without loras. This >>25817 for example, needed 5+ specific artist loras that individually play ball with dragon girl concepts (no good dragon girl lora around). Most other loras either made the dragon girl more human, furry, and in the worst case caused extreme frying (several cases). And these are my "decent" loras. So always be mindful of any new untested lora you download.
>>25819 >>25820 thanks bud, i really appreciate all the help i get in this thread, its helped me understand the fucking planetary sized nuance behind using Pony when compared to anything else stable diffusion. I was curious because in all my testing since sunday, it really just seemed like using MOST loras i've grabbed so far TANKS the quality of anatomy, rarely does it ever fully break anatomical accuracy, mostly fingers and toes, placing nails underneath fingers and toes, that sort of thing. It always seemed like a good rule of thumb to keep characters at strength 0.75, then add strength to the character name/trigger itself, but in Pony, nine times of ten, adding any kind of strength outside the LORA at 0.75 will completely break the accuracy of the rest of the prompt. Most LORAs seem sensitive to this ,were compared to a non lora prompt, adding strengths works as expected. Funny stuff. Only makes me want to become an XL/Pony LORA baker even more, there's a huge demand for it and maybe a handful of people on civitai can do it consistently well.
>>25821 > it really just seemed like using MOST loras i've grabbed so far TANKS the quality of anatomy, rarely does it ever fully break anatomical accuracy, mostly fingers and toes, placing nails underneath fingers and toes, that sort of thing. Yes it's all highly variable and unpredictable. A lot of loras are like water and oil. I don't particularly use a lot of character loras, mostly mix styles together. But a trick you can do is use an already known, stable and flexible lora as a stabilizer at partial weight, if it's similar enough to the other ones you're using. This is also not a magic solution if the lora you're trying to stabilize is complete garbage. The only tip I can offer is to check the posted images (by creator and user submissions as well), see if it has a variety of poses, if it can handle hardcore, oral, and basically anything you intend to use it for. Notice also how much the style/character becomes diluted under stress. And if you find a decent baker, you may want to check what other stuff they've made as well. >Only makes me want to become an XL/Pony LORA baker even more, there's a huge demand for it and maybe a handful of people on civitai can do it consistently well. It's another rabbit hole, but go for it if you have the free time. Properly made, flexible loras are scarce.
(892.69 KB 1024x1024 00020-2844217807.png)

(924.62 KB 1024x1024 00023-1496315435.png)

(1.18 MB 944x1216 00046-1519304419.png)

(1.45 MB 1024x1024 00050-3201267619.png)

>>25822 shitty LORAs make me genuinely sad, its like finding a concept i really want, then watching it bastardized and wasting my time looking like picrels. Now i want to make: Clown loras, Mime loras, like 20+ different specific characters, iterate more on that PC98 Lora that Anon made to sway it another direction, and ball jointed dolls, which is a fetish ive had for ages. these BJD results makes my heart break.
>>25823 *posting this only just made me realize that the example pics posted by the LORA maker are flat out using the wrong prompts, i need to actually sit and laser focus these prompts before i just drop them into forge and act surprised when they're shit, even when they come from the lora baker kek >source_anime >on a photorealism model
>>25823 jesus christ
(1.22 MB 1024x1024 00053-582118680.png)

(1.29 MB 1024x1024 00054-3125133341.png)

(1.29 MB 1024x1024 00057-3586643788.png)

(1.17 MB 1072x1072 00065-2741563027.png)

>>25825 >2 bjd models on civitai >NEITHER are good theres a big vein going 3d on my head right now. why did this guy fucking train on cybernetic LED-eyes robots for a BJD model?
fuck it, my autism has been spiked to its nuclear meltdown maximum, i'm using my 500 civitai credits to make a bjd lora. Pony.
>>25826 Stuff like that is fine as long as it's tagged properly. Most problems come from poor tagging, summoning things without asking for it
(1.94 MB 2816x2304 00218-2999521449.jpg)

>When the gen screws up and makes spreading with clawed feet instead of hands. I'm having fun with this. I found the trick to stabilizing dragon scales is keyword blending. Trick the thing into thinking it's creating long gloves and thighhighs on the first 20% steps, then swapp for nude and scales. Reverse outfit works too, and I suppose this extends to other monster girls. Surprisingly effective cheeky method (deleted last post, compositing error)
>>25828 Yeah i keep forgetting tags are such a crucial part too. Certainly a good thing to be reminded of as im on the journey of building the dataset kek
>randomly got style I liked (1st pic) >if I remove "white micro bikini, armpits" that was between artist tags, the style changes drastically >can't get the desired style even with padding tags with nonsense/spaces getting consistent style on nai really is annoying.
>>25832 It's actually a general problem. Artist tags are not guaranteed to be any more stable than loras, and their weight is influenced by concepts. All concepts have style bias and vice versa, because the distinction between concept and style is fuzzy for the AI and the data provided is asymmetric. (Artist1 has 50 pantyshots, Artist2 has 10 pantyshots => pantyshot will have a bias reflect Artist1's style) Now extrapolate this logic to a fucking clusterfuck web of a model.
>>25833 I have to wonder if these A.I models have noticed by now how massively skewed certain degenerate fetishes are to certain regional styles, western for inflation/living doll/etc. I mean, I just started noticing this shit yesterday while browsing citivai. "Like 99% of the western-adjacent styles im interested in are filled with inflation and living doll examples but none of the eastern ones are wtf" lol
>>25834 Well it clearly has a source bias. Those subjects are more commonly done by western/furry artists. It follows that lora sampling will draw from a western pool and thus result on a western style bias.
>if these A.I models have noticed i mean, either the weights reflect it or they don't
>>25835 yeah thats ultimately what i mean >>25836 quote me next time pussy biiitch
>>25833 Yeah, "micro bikini" changes the style regardless of the tag placement(seems to amplify fizrot). I don't think I've ever encountered it being this bad tbh And it's not like fizrot has a lot of micro bikini pics, there's only 4 of them on dan (fkey has none though).
the thread gets like eight posts a day, it's not hard to figure out who's talking to who
>>25839 >eight posts a day all me
>>25840 wow general hentai diffusion himself, it's an honor to meet you sir
>>25841 ligma balls
>>25839 It's comfy. no spam, can actually have discussions. Bit more traffic wouldn't hurt but don't wanna alert the retards
>>25799 I remember some old discussion about forge couple, some anon stated that character loras better working when placed right in the lines of characters positions. Very neat extension though, super easy to make some extra uoh, hope it will be transferred to a1111 or regional prompting will have some 'easy mode' or smth
(1.23 MB 1024x1024 00172-3102992596.png)

>finally figure out how to tardwrangle the LORA >sure its uncanny, creepy, and my dick can't get up looking at it but its interesting >what the FUCK is that in the background?
>>25844 Catbox? There's no metadata on these.
>>25845 That would be anniki
>>25844 They're not lines, it works like this >>25803 If your characters fall right inbetween those 2 zones it busts the gen, basically.
>>25845 bro spoiler that shit plx
>>25847 >he's watching over me while i fail miserably to work this lora out thank you big bro >>25849 no, bust your nuts after you save whatever posts you like.
>>25846 https://files.catbox.moe/lly8m5.png https://files.catbox.moe/binovs.png https://files.catbox.moe/epu0vc.png >>25848 So it's just better to put character loras in the global effect line too, or that doesn't matter at all where to put lora in the prompt?
>>25851 Always global, I don't think loras work by line, I could be wrong though If it's tag triggered as most are, global lora, tags on specific line should do the trick
>>25845 that's a picture of me looking at this picture
(1.85 MB 1280x1408 00455-2245356298.png)

(1.79 MB 1280x1408 00451-3838172442.png)

(1.99 MB 1280x1408 00446-3514497420.png)

(1.81 MB 1152x1408 00431-357558615.png)

Well I've gotten it relatively consistent now on all poses, sex and whatever else. Challenge beaten. Some tips: In extreme cases where your char description is extensive you wanna reduce bleeding into other characters as much as possible, it's always an option to avoid 1boy or multiple X tags (for background characters). sex w/o 1boy will lean towards pov/crotch/half out of frame, I think it understands them as props and doesn't apply tags to them (as frequently). Also if you didn't know: https://donmai.moe/wiki_pages/tag_groups That's a comprehensive index of (most) working tags by category. Useful for background stuff https://donmai.moe/wiki_pages/tag_group%3Alocations Sex poses https://donmai.moe/wiki_pages/tag_group%3Asexual_positions Angle control https://donmai.moe/wiki_pages/tag_group%3Aimage_composition Sex acts, clothing, hairstyles, objects, and basically everything. Good for checking obscure shit you weren't aware even existed. Most is valid, some has either to little data to work or wasn't taken into account by pony, but it's the exception.
Im finding even Pony is just bad at everything i want it to do, either because of bad LORAs or its flat out not trained in what i want to say i'm disappointed, let down, violently sad, is an understatement.
>>25856 Could be a skill issue if you're just starting. What do you want to do exactly?
>>25857 man i was flying high since sunday but it just seems like LORAs and the lack of training on specific """niches""" are its achilles heel. Genuine capability, but no shit, because it's heavily trained and specialized to do specific things. People make it sound like a one size fits all, it's reaaallly what one anon said in this thread, down to the myriad of settings and potential issues brought on by training/lora. In this case, the most basic thing imaginable: GOBLIN GIRLS. It's doing "okay" but heavily skewed to specific western styles, and doing anything 3D is insanely tough, even with decent LORAs honestly could more than likely be a combination of my skill issue going into more new territory + lora problems. agh.
(2.22 MB 2400x2720 5.jpg)

>>25858 I'm that anon. What are you struggling with, wonky anatomy? lora combination frying the gen? wrong color gobbos? Here's one rather basic trick. If you're getting western styles, neg: source_pony, source_furry, source_western But remember, negative prompting will limit the data pool, meaning you pay a composition "tax", if you run out of valid combinations due to increased complexity, your anatomy will start fucking itself all over the place. Few negs are okay though.
>>25859 I'm calming down now, i just realized most pony-derived models are trash, it's not a pony issue. Trying out Aoi Realistic Pony, really simple prompting as usual so it's no prompt issue, just.. model. No loras. I'll have to try that source_western from now on, i forgot to even consider it. My whole problem with that model came down to the fact that it NEEDS a specific resolution, must always be 768x1280/1280x760, anything off or even base sdxl resolution will destroy gen quality. Pony isn't even this picky.
>>25860 >came down to the fact that it NEEDS a specific resolution, must always be 768x1280/1280x760 That's odd. Yeah, pony and autism can handle pretty much anything, although one should always use multiples of 64 due to some incompatibilities (like with Forge Couple) and other tools freaking out on your gens. >i just realized most pony-derived models are trash, it's not a pony issue. Trying out Aoi Realistic Pony, Yeah, I'd ditch most pony-derived models if you're trying to go beyond 1girl tier prompts. Some are neat but they're specialized and struggle a lot with anything weird you throw at them. I'd stick with amix-pony amix-dpo and let loras handle the rest. Do you want 3D or 2.5D? I can recommend a couple of old loras for 2.5. They seemed to blend ok as far as I recall
>>25861 Thanks, gonna clear my folder of anything non autism/normal pony. i wanted to do 3d just to see how XL models handle blending reality with a fantasy concept, at least to get some creative inspo to get back into blender. Can they really just not handle it or do i have to reeally prompt it right? Wouldnt mind trying 2.5d i just see it everywhere and none of the style variations appeal to me.
>>25862 >i wanted to do 3d just to see how XL models handle blending reality with a fantasy concept I see... Yeah that's tough. Maybe through some deep learning shenanigan, but I don't think there's a huge direct source of 3D hyperrealistic fantasy porn. >Can they really just not handle it or do i have to reeally prompt it right? I mean, maybe they can, but it's complicated because when you enter "goblin" the model is likely not gonna have any off the bat 3D goblin data, it has a bias towards the goblins it knows, which aren't 3D, so it'll make it less 3D. But maybe if you really really hammer in you want 3D? Like neg 2D, add realistic, etc. Pony isn't really made for 3D though. Here are the 2 loras I found OK to blend Very generic one, but seems solid https://civitai.com/models/15792/neoartcore-lora And this which I've mostly used in the past as a skindentation enhancer https://civitai.com/models/316439/gpthicc-style-xl-pony-xl Mind you I've always used these at low weight, as style mix tweaks, so I can't really say how they'll blend on high weight with other stuff.
>>25863 Oh and also, I have NO clue if they'll play ball with goblins or not. Sometimes a lora is mostly fine, but you they have "pet peeves", like an obscure tag that completely fucks them up for some reason. Dunno, what I do is simply lock the seed, swap a lora, gen, swap, gen, swap, gen, swap gen, Rename your outputs to the lora name and by the end of it see which improve the gen. Use the best or combine at different weights to shape something new. That's about all you can work with
>>25863 >>25864 I guess im thinking too much like this shit is black magic that can 2D>3D with little issue, i'll give lora mixing a shot or somethin' in that case.
>>25865 Have you tried using the refiner and swapping from base pony/tismmix to one of the realistic merges early on? Or genning a 2d pic and running it through img2img with a realistic model, if you do it at low denoise you could run it through a few times gaining realism each time.
>>25866 oh i have no fuckin clue how i would achieve either of those things, still havent looked into how refiner works and img2img was fun to play with for a bit but tough to get a grip on. i wanted to try refiner for the doll experiments earlier, hoping thatd fix the anatomy right up.
>>25867 >>25864 here Do you want realism realism or CGI esque stuff? I'll run some tests in lora mixing if you want, tell me precisely what you're after though. I like a challenge
>>25868 the image in my head basically went: "I want it to look like i just discovered this homeless, dirty, near-feral shortstack goblin girl at a local park, then brought her home, gave her a shower, fucked her a few times, and now she's wife-material." So, realistic, no disney or 2.5d. Kek. if that's honestly out of the realm of pony i understand.
>>25869 Yeah I don't even think I've ever seen a realistic goblin anywhere ever. I'd go as far as to say it's out of the realm of what Stable Diffusion can do (right now) If you're really committed you could try what >>25866 suggested, but from experience i2i without base data tends to murder quality. For example, if it doesn't know what a big goblin ear is, it will find the closest thing resembling to it, as for everything else, it'll become more human with each pass. Controlnet is also an option I guess, although I'm unsure what particular one would be best for that project.
>>25872 GHOST CUM DICK GHOST CUM DICK GHOST CUM DICK
>>25871 Have you tried everclear and this https://civitai.com/models/428826/damn-ponyxl-realistic-model Last one seems to have some fantasy concepts. Maybe it can handle a short green bitch with pointy ears?
>>25874 Holy shit that model is CRACKED lol'ing at the variety of degeneracy and brainrot, and it looks pretty consistently good. Like maybe one or two gens here and there have some odd finger additions but nothing extreme. I'm gonna give it a shot. >that f*male gator
>>25875 Yeah that's why I suggested it. You could also probably do an image search for realistic gobbos and then check the resources used. Can't check rn, but there's a chance
>>25876 DAMN wins so far, it got the idea first try and the only problem i'm having is really grainy hands and feet, but the anatomy isn't wonky and she looks very goblin. prompt strengths are overly strong, prompting a big belly and tits gives her SSBBWW or some crazy fetishy shit.
>>25876 >>25868 uploaded some cherrypicked of today's fails, apparently i have almost 300 of them. theyre noteworthy for lora mixing/no lora just rawdogging a model's capabilities/trying negs (i think?) and crazy prompt experimentation this is certainly a nice combo of challenging gens to try for https://catbox.moe/c/n7mrij
damn that's a lot of posts for this place
>>25872 Spoilered image is the best one.
>>25879 mostly me
>>25881 tan brats in white clothing flashing pink nipples ToT
i think i found a new autistic fixation, trying random schizoid civitai prompts and testing new model prompt adherence through it very interesting stuff.. >view from side, booty, wet and dripping with ky jelly, body completely drenched and dripping with honey and liquids, sitting on the floor with legs wide open, small chest, dynamic posing, moaning, ultra-detailed facial features, burkinabe woman, sexy plump bbw thick curvy plussize light skinned black woman, proportionate bbw body, cute, hyperdetailed, huge ass, wide hips, stretch marks on hips, perfect big butt, plus size ass, thick chubby legs, full makeup, naked, being slutty dripping thick honey, completely drenched in honey, completely covered in honey, sultry expression, seductive, hyperdetailed, 32k uhd, extra sharp focus and details, 100 megapixel highly detailed,
>>25880 because that's the specific drawing that it's aping, lol
>>25883 they are still the cutest in micro bikinis
>>25886 funniest, yes. They're cutest in oversized clothing that has too many layers and frustrates them because they can't seduce ojii-sans
>>25888 peak SEXO
@BJD anon I'll bake you a LORA if you can give me dataset material
>>25890 sure, itll be a while though. Finding consistently decent pictures of these things is surprisingly tough. I gave up when i realized i maybe had like 5 good images in my dataset kek. Lotta watermarks/censor emojis/124px compressed jpegs/etc..
swear on me mudda i'm never gonna fuck with negs in any XL/pony model, if a LORA suggests negatives, its a shit LORA. simple as.
(2.40 MB 1664x2288 catbox_kilcxz.png)

(2.77 MB 1664x2288 catbox_z8pc79.png)

(3.30 MB 1664x2288 catbox_sgdi75.png)

(3.05 MB 1664x2288 catbox_0tlsaq.png)

>>25888 box for rin?
>>25813 got my 3090 in and it does about 2.6 it/s on the same settings definitely more card than i needed in retrospect but i'm happy to have it
>>25896 >3090 download mixtral instruct 3.7bpw exl2 and enjoy slippery sloping down the LLM rabbit hole
>>25896 and more importantly, it can do 1 it/s while i'm playing video games >>25897 i'm looking forward to seeing what local models can do, claude opus was fun but sonnet doesn't have the horsepower for what i wanted
>>25898 mole
is image interrogation meant to be insanely shit? >takes forever to load the model >finally loads >cant caption to save its life tried two models and had to reload both times because neither could unload properly..
>>25901 its the "ffs I missclicked that button again time to wait a minute" button
>almost 1200 someone make a new thread soon please thanks
(29.59 KB 474x315 bread.jpg)

>>25903 Bread baked: >>25904 >>25904
>>25900 Box for the tianliang-looking one pls


Forms
Delete
Report
Quick Reply