Wife belly

https://exhentai.org/g/2939153/ce9ee2d96a/ Is it possible to do this pose with Pony? What prompts would I use?

bread is there a forge plugin for better prompt blending? having to autistically stick BREAKs in everywhere is getting old. >>25908 M legs, arms behind back, on back

>>25908 spread legs as well

>>25910 Nice, can you catbox the image so I can use it as an example? I always have trouble with it

>>25911 metadata button ignore the wall of shit at the bottom, that's just latentmodifier

>>25912 oh you're right it is there my bad

What was the name of extension to store gen metadata inside? >>25907 https://files.catbox.moe/aowpwm.png

>>25914 isnt it in the OP the hdg script

>>25915 I know about catbox, but there was some extension for automatic

>>25916 Shouldn't metadata be automatically stored in the picture if you're using forge/A111?

>>25917 yes, but like 4chan, 8chan strips metadata when you upload the images also unrelated, huohuo uploaded, outfit is a bit jank, but it's really inconsistent in the art https://mega.nz/folder/OMY2RRxQ#zYaokRZV361WSNImfrBCrQ/folder/KUhQFBCA

>>25917 Yes, but it will be purged when you uploading it here, for example, that extension was storing it in some finicky way to prevent that, just like in this post >>25910

>>25919 i think forge just does it automatically

>>25920 I found it, its stealth png info

>>25921 yeah, that still works, but we have a script we can use to grab the metadata directly in the browser. just a nice quality of life

Is AMD viable yet?

>>25923 no and never will

Does anyone has an idea to work with huber loss? I could maybe give it a try, my trainings have been getting kinda lots of unstable loss peaks, it can easily go from 0.077 to 0.093 and I've heard huber could help with this

>>25923 sure, get the zluda implementation

>>25925 Check out all of my lora, all of them use Huber loss You can find the tomls at the link on >>25918 On another note though, an unstable loss is not really that surprising with SDXL models, because it's just... jank

>>25923 Yesterday at AMD's conference, it was announced MI35X series, which is supposed to outperform Nvidia's H200. So basically, if you're not a millionaire with unlimited resources, this information is useless for you. Nvidia is still king for consumer tier, and it seems it will be like this for quite some time, since ROCm doesn't seems to be interested in adding more support for older consumer tier gpus. Hell, not even for the 6000 series.

>>25927 Oh I see, alright thanks, gonna give it a look >On another note though, an unstable loss is not really that surprising with SDXL models, because it's just... jank Why does this not surprises me... sigh. I hate SDXL

https://civitai.com/models/492306/encorewutheringwaves-for-pony?modelVersionId=547379 This is so toasty compared to the lora some anon from here made The civitai dude probably trained it on game screenshots and called it a day

is it possible to gen on a laptop without murdering it SDXL seems like its really demanding

>>25930 Yeah, I know my encore lora isn't entirely consistent on the outfit, but at least it's not baked like that one! I'm fairly happy with it tbh, because encore turned out flexible enough to put her in a whole bunch of different outfits too. Same with verina, verina was actually harder because she had like 12 pieces of art in total. Regardless, so long as people enjoy using my lora, it's worth the effort to make them

>>25929 Yeah, same tbh, sdxl is just not a great model

>>25931 I believe forge is able to run sdxl on 4gb of vram, it'll be slow, and you probably won't be able to upscale very far if at all, but it's possible. How much vram do you have?

>>25934 I'm planning to buy a new laptop soon, was just wondering if its even a good idea to run SDXL on it or just stick to online services like colab and tensor to not murder the thing. Probably a laptop 4050/4060 most likely for reference

>>25935 4060 8gb will be fine for sdxl but not future-proof for further models. I was looking for a laptop upgrade myself some time ago and was a bit surprised how much 12gb+ vram models cost. And if your laptop doesn't have absolute bottom-tier assembly you shouldn't murder it with genning, no. Undervolt your gpu though and have hwinfo to check temps if you're paranoid.

>>25936 What are some good laptops that have good build quality and don't murder my wallet? Framework seemed like the best but I really don't want to spend upwards of 2k on a laptop.

>>25937 Sorry, don't know what's good now. I wouldn't take HP and noname chinese brands, that's all I can say. I've been using Dell G5 laptop for like 5 years already and it's been pretty solid except for battery life. Mostly bought it due to lack of cringe gamer design though.

>>25937 doni't be a retard and just buy a stationary pc

>>25939 Some of us need the mobility a laptop offers else yeah a desktop PC will always be better.

>>25935 honestly now is one of the worst times ever to buy a gpu, we're in a point were the future is AI, and they're making sure to jew the fuck out of consumer GPUs as possible by crippling VRAM capacity do NOT buy anything less than 12-16 gb if you ABSOLUTELY need a gpu. You're going to deeply regret getting 8gb the second you have it in your hands. 8gb is the absolute bare minimum for all things A.I alone.

haven't checked in since the usable SDXL pony models started getting used. what's the go to these days for prompting and training? i miss simpler times with basedmix and aom2

>>25942 autismmix confetti for prompting and pony v6 for training, afaik

>>25940 literally what for, in that case you just get a shitty laptop that does the bare minimum

>>25942 I wish I asked the guy who trained a 1girl lora for me to do it on animagine just so i can use the inpainting model, way more useful 4me and now im cucked. otherwise yeah pony

>>25944 nta but Im rentcucking in different places and planning to move to another country. I'm gonna have to sell my pc that is at my parents' place, probably gonna have no use for it

hook uploaded: https://mega.nz/folder/OMY2RRxQ#zYaokRZV361WSNImfrBCrQ/folder/jNJ2yBhJ fun fact, hook only has like 100 pieces of art despite being a launch unit for hsr, most of which is completely unusable for training. as such, this dataset only had 10 images in it total. I used masked training to try and handle there being such a small dataset, and it seemed to work out well. I didn't try to put her into other outfits this time around, so can't say anything about that, but it probably would work well enough. takes to style lora very well

>>25947 high quality. is that really all there is to the prompt? do you drop the negative for hirezfix?

>>25949 yeah, that's it, I don't use negatives in my gens, I let my lora do the work. same with not using the score tags, source tags, or any other quality tag for that matter

>>25950 that's pretty neat

>>25948 Blessed taste, Mitsudomoe is kino. More Mitsuba is always welcome, though I think her belly pudge is a little overexaggerated there.

>>25953 Maybe a little but also not too far off

The major is recuperating after an injury in the field. Please be sure to carefully guard her body while she rests.

>>25955 hot

>>25947 HSR mostly releases adult women or men so not surprised she doesn't have a lot of art. Being a lolicon in a hoyo game is suffering sometimes.

>>25957 agreed, though hook is an unusually small amount compared to the other loli characters i've made lora of.

>>25959 She's sure happy to be outside.

>>25962 beach brats with popsicles, one of my many weaknesses..

Seeing someone else using it has gotten me interested in prompt padding (basically just fill to max prompt size with quality tags, make sure neg and pos prompts are both full) again. I used to do it a lot in local but never really experimented with it in NAI. If you remember the old schizo negatives, that was a type of early shitty prompt padding. Anyway, has anyone else experimented with padding and have any ideas on ways to do it better (whether NAI or local)? It does seem to make some difference. catbox for aco rwby stuff, going to do more anime style and content testing later. https://files.catbox.moe/ghwut3.png (embed) https://files.catbox.moe/pfyna1.png (embed) https://files.catbox.moe/t8fl5c.png (embed) https://files.catbox.moe/gy4biw.png (embed) Y'all are more technically minded than /h/ these days, anyone tried this kind of thing?

Any of you script wizards know how I could gen on colab with loras? Been using this colab for a while https://colab.research.google.com/github/R3gm/SD_diffusers_interactive/blob/main/Stable_diffusion_interactive_notebook.ipynb?authuser=2#scrollTo=eHO0LSC--4WF but it's old and hasn't been updated in ages so some SDXL loras do not work in this and I don't have the technical knowledge to make them work Not particularly expecting anyone to but would be extremely helpful if someone figured it out since not everyone has a great rig to gen stuff on

yuuki haru: https://mega.nz/folder/OMY2RRxQ#zYaokRZV361WSNImfrBCrQ/folder/3YxBEALA

Fat

>>25969 tastefully thick. squeezable.

>WARNING[XFORMERS]: xFormers can't load C++/CUDA extensions. xFormers was built for: >PyTorch 2.3.0+cu118 with CUDA 1108 (you have 2.3.1+cpu) I want to die. >>25955 Nice.

>>25970 She still needs a little bit of cardio correction.

>>25972 Does plapping count as cardio?

>>25973 Definitely

>>25975 Very nice

drs style: https://mega.nz/folder/OMY2RRxQ#zYaokRZV361WSNImfrBCrQ/folder/mBZTRBbS

>>25975 Box please

>>25976 thanks! >>25978 Meta's there as other anon pointed out if you don't want to use browser extension for whatever reason, you can use this https://github.com/neggles/sd-webui-stealth-pnginfo I'll still catbox the base gen though. https://files.catbox.moe/pgntq3.png

>>25977 Any updates about adding this to easy scripts? >>25745

>>25981 oh right, forgot about that, been busy so I sort of forgot, I can update it today quickly

>>25982 update pushed to dev, and I gotta say, absolutely 100% fuck that arg, it just had to clash with dora_wd, which made an easy change into a hard one.

>>25983 I don't know why kohaku just decided to make perf worse, not even documenting that. How is that conflicts with dora?

>>25984 the new arg explicitly crashes the process, so I had to add a new limitation to dora, which already has a ton of limits to it, aka, must be locon, loha, or lokr, cannot be used with network dropout, and cannot be used with bypass mode, which isn't locked behind locon, loha, and lokr only. as such, you can get into a situation where you had dora checked, then you swap to like ia3, for example, then check bypass mode, which would screw up all of the logic on what is enabled or not.

>>25985 Alright, ty, why dora is so restricted for features btw? I still dont get it, like half the training speed on 4090, seen some complaints on 4cuck that 8gb vram got filtered out too at inference time using it, no dropout (well this at least understandable by design of that algorithm) and results is comparable to just adding convolutions? Maybe I just trained it wrong, but it's just don't differs from locon this much

>>25986 I pretty much only train dora, and I have zero issues with it. the main draw of it, for me at least, is that it generalizes much better

>>25987 Well, I know some prompts that when you trying to gen it, it will just overweight loras for some reason, for example https://files.catbox.moe/3h9o5v.png I tested like everything, longer run, higher lr(this will just blow everything and output almost indistinguishable images when prompting something from dataset and still won't get enough of the style generalized), prodigy, dora, but only solution to get it more tianliang resemblance was to train from autism and overfit, but not at the fast pace, it starts to break fingers on 1 weight and almost unusable with full weight when prompting from dataset or not using anything else that will try to overweight it. Wonder if there is some better way to do it, truly, dora was a bit better, like a second place for that, but still on par with everything else, not even close enough. Besides all that, this prompt will do just absolutely fine if artist just have some ass focusus, asses in swimsuits, bluearc, etc. Overfit one to hell, extend dataset for most popular cases with that and train again, lol?

>>25988 I mean, it pretty much boils down to pony, i'm pretty sure. after all the hashes break the entire TE. I'm gonna be honest though, my tianliang dora does pretty damn well, and that one is one of my older ones at this point

>>25989 This one? https://mega.nz/folder/OMY2RRxQ#zYaokRZV361WSNImfrBCrQ/folder/fQoziYrS It's exactly what I've tried to nail down, it just feels like "not enough" especially when comparing it like that. Slight divergence from the dataset and for some reason it's all falling apart like a house of cards, although there seems to be not a big difference with cunnys, and I'd prefer the left one myself.

>>25990 yeah, i'm talking about that one, but also you don't want to use a negative or score tags with it

>>25991 when in doubt upweighting works, in order, 1, 1.2, 1.5 + extra loli shiroko

>>25991 Okay, just did it and this "lol" printed shirt was generated >>25992 Yeah, it's another way to force the style more on autism

>>25993 apparently it still handles 1.7 fairly well too

>>25995 cute

>>25993 why would you wear a shirt that says "lol" wtf is wrong with zoomers

>>25999 they should wear "loli" shirts

so you probably have seen how fucking bad sd3 is with anatomy and females in general nonetheless someone on cuckchan managed to gen literal csam shit with it >>8063010 (4chan hdg) How fucking incompetent is sai?

>>26001 haha oopsie poopsie we were so busy censoring all the stuff that make the fat childless shrews in our social media department upset (beautiful women) that we forgot to care about the illegal shit, what a widdle fucky wucky xP

Fat bunny Friday

>>26003 now thats a chubby bunny (challenge)

>>25990 i borrowed your TOML and i'm getting 600 s/it on a 3090 and a wall of triton errors, is that normal

>>26000 True >>26006 It's not my config, if you took one from mega link, it's this guy config >>25996 But triton errors is a common thing on windows, you can ignore it, 600s/it is not something good though

>>26007 cute and based

>>26006 Triton errors are normal, make sure you have checkpointing on, it uses 42gb of vram without it, with, it's only like 12 though

>>26009 Based fellow undressing enjoyer

>>26011 bunny girls are looking great. I'll post some more undressing in a bit

Playing around with DoRAs. Here's Koishi Chikasa. Catbox is down but the previews in the MEGA have metadata. Thank you to Mr. John Easyscripts Anon as usual. Highly recommended to use loli or plump/chubby tags for this one. https://mega.nz/folder/OoYWzR6L#psN69wnC2ljJ9OQS2FDHoQ/folder/upxl1ShJ

okay what the fuck euler a is actually goated for sdxl (second two images are shota and kemono respectively)

>>26015 yeah, i've pretty much only been using euler a for SDXL, don't see much reason changing it when I get good results

>>26016 oof, accidentally uploaded the fucked up one

>>26017 uoooh

healthyman: https://mega.nz/folder/OMY2RRxQ#zYaokRZV361WSNImfrBCrQ/folder/bEhzSbBb

>>26019 This one looks like it could be fun to mix with, thanks.

>>26010 thanks, got it running the three errors i was hitting were >going OOM and swapping to RAM turned checkpointing on >bad file paths because i'm on windows opened the TOML with a text editor and find-replaced \ with / >avr_loss disappearing after a few minutes set no half vae

>>26021 the average loss dissapearing is just because your terminal window is too small, kohaku shoves so much shit in one line of info that some of it just gets cut off, and you can't change it once it's started btw

dooge: https://mega.nz/folder/OMY2RRxQ#zYaokRZV361WSNImfrBCrQ/folder/qAhkmCBR

thank you mr. john easyscripts for the antifreeze dora it's fucking fantastic when pony decides to actually generate the foot properly

>>26013 sexo

>>26024 first of all, "mr. john easyscripts" is probably one of the funniest things i've heard, and second, glad you like it! even though i'm fairly quick with my lora, I do try and put a lot of effort into them

>>26022 i see i see two failures and i'm running my third now >changed batch to 8 because batch 2 was only like 30% of my vram and didn't reduce my learning rate, charbroiled >reduced learning rate but didn't toggle the setting to be able to resume, underbaked photoshopped all the watermarks and scan artifacts out, next plan of attack is to retag everything again because i thought 75 tokens was a hard limit the first time around this is sort of fun

>>26019 First time tried using your dora in comfy, my VRAM usage jumped from 6.8GB without dora to 12.5GB with it. Is this normal or should I be using some custom nodes to alleviate this? Currently using default lora loader. I have 11GB so naturally gen times plummet to unusable speeds...

>>26028 Apparently I needed to update comfy with dependencies. Now it takes 8.3GB with dora the first time and for some weird reason 7.3GB the second time and onwards (or if I just cancel the first gen).

end result: all of them kind of blow

>>26028 I am almost 100% sure comfy did some change with the dora implementation a few weeks back that made it significantly worse, because on forge and auto the dora I make work very well.

>>26033 Wasn't realism one of the things this board has banned? Yeah, it's the thing talked about the most in the pinned post.

>>26034 oops my bad

Settle down, shut up, and cook.

>>26036 What did he mean by this?

cooking kind of sucks actually

shiratamaco: https://mega.nz/folder/OMY2RRxQ#zYaokRZV361WSNImfrBCrQ/folder/TUpjQZTD

>>26039 Cute, thanks.

i request artist loras of these two artist pretty please me vramlet https://www.pixiv.net/en/users/10199514 https://gelbooru.com/index.php?page=wiki&s=&s=view&id=7217

Sloppa but I will always love cunny.

>>26040 >unlolified

>>26041 ikuse looks really nice, I'll take a stab at it later.

>>26043 Yes, I have sinned.

>>26042 Good cunnysloppa.

>>26045 I will repent. >>26046 good beach brat sloppa

>doing tests >color and style abruptly change for no obvious reason it's got some life though, getting somewhere >>26040 >>26046 >>26047 nice

oh no

>>26049 that smug smile... she totally manufactured the situation, damn brat

>>26050 Yes, she's a total whore.

>>26049 An unexpected twist!

i'm still not fully happy with it but at least i can get to the part that matters, endless lora mixing

>>26053 Cute Olli

>>26049 now post the aftermath

WIP for the fella who requested an ikuse LoRA, still experimenting with parameters and DoRAs learn faster than I thought so the first run came out overfit.

>>25961 MORE

https://mega.nz/folder/OoYWzR6L#psN69wnC2ljJ9OQS2FDHoQ/folder/ygQBQZra ikuse for >>26041 Also included the toml which is based on my old LoCON settings and John Easyscript's DoRA settings, should work quite well for 16GB+ (possibly 12?).

>>26058 LESGOOOOOOOOOOOO Thank you bro it's very nice

>>26061 Cute! Could you share a catbox for the second one? I like the style

>>26062 https://files.catbox.moe/6ou5tu.png Metadata should be available with stealthpnginfo but sure. Sadly this style doesn't hold up when used without character/copyright tags, but I really liked how it worked out for Illya/Kuro.

>>26063 Thank you! The metadata was indeed available I was just being a dum dum

>>26055 I will, when I find some time to gen. >>26060 She's a cutie. >>26061 >>26063 stupid sexy Illya.

has Pony/Autism with LORAs caught up to NAI yet? Tired of paying for a model that gets left behind when it comes to waifus.

>>26066 I'd say yes. There are caveats, mostly with style, but certainly worth trying if you're missing your waifus.

>>26067 NAI is really good at touhou/gacha/vtubers. but those can get tiring after a while. doesn't help that it's at the mercy of Gelbooru tags.

Danbooru*

>>26066 ehh i've seen a few posts in /h/ that tricked me but they were probably a couple hours of effort each

>>26068 It can do quite a lot of charas, not only the most popular ones though. But yes you're missing out on the latest or obscure waifus. >doesn't help that it's at the mercy of Danbooru tags True, you technically can use their furry model as a crutch (I did test it) for some concepts that anime model shits itself with, but if you got a good gpu at your disposal just go play with local.

>>26071 my 3060 ti needs an upgrade soon, but i can bum it out with Forge.

>>26066 Pony is super scuffed but you can certainly get better results than NAI with certain subjects (and of course, training styles/subjects NAI doesn't know). The major caveats in my experience are the "hashed" tokens causing jank and the noise offset is SUPER scuffed, which you can observe comparing Pony outputs to NAIv3 and Fluffyrock prompting for dark scenes. That might be fixable with a LoRA or LoRA merged model, not too sure.

>>26066 On style yes, if you don't like the default pony style you can just plaster it in loras or merge or retrain it. On promptability fuck no, the text encoder is fucking garbage.

This stuff looks really good. Any ideas what they're using or how they're prompting? www.pixiv.net/en/users/90281477 www.pixiv.net/en/users/105818043

>>26075 looks like your stereotypical nai gens to me, likely just using the same artists as nearly everybody else

Requesting a Style LoRA for this artist https://twitter.com/ceramic06 Very cute and funny stuff, also I like his linearts

So Forge is officially kill for normal use? https://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/801

>>26078 yep I'd say if you don't miss any features stay on the good commit for now, it still performs better especially if you are 6-8gb vramlet.

>>26065 I found some time.

>>26076 Do you know what artists? I can see Modare and Rei in there

>>26081 no, I don't, I usually am quick to forget most artists

>>26083 Very nice

>>26066 lol

>>26083 >>26085 cute >>26084 extremely cute

fuck man came back after several month hiatus to find hdg in even bigger disarray than it was, that's sad

>>26088 Yes, I have basically stopped browsing the threads.

suri_solari: https://mega.nz/folder/OMY2RRxQ#zYaokRZV361WSNImfrBCrQ

4chin hdg has been garbage for a while. I don't think cadcade or wdV or any new local model being good can save it at this point

>>26088 That's what bot posting, discord raids, local/nai drama fagging and model shillwars will do to a thread that receives little to no moderation. Whoever is in charge of /h/ really doesn't like being stuck with the general and it clearly shows. I also doubt they will create an AI board, despite there being so many generals strewn across the site, since I suspect they just don't want to paint a clear target on themselves in regards to future legal shit. I'm glad this general is still around though.

I think it's also people not caring about AI as much aymore, novelty's wore off. So shitter and schizo majority remains

>>26093 Don't think so. Hypefags don't particularly bring quality with them. With hdg it's just the case of literally 2-3 schizos raping the general with the help of malicious moderation.

Hi, I'm new here. Can I join?

>>26095 Howdy and you are more than welcome to join us here.

happy fourth of july, everyone

>>26095 It's pretty dead here but sure newfrens are welcome.

>>26097 cute patriotic beach brat

damn brat...

>>26096 >>26098 Thanks It's always a tough choice to play or train/degenerate

>>26101 Yeah sure is but I have some time off work now so I'll be posting more often for a bit now.

I've almost exclusively been using imagegen to make pics for textgen character cards lately. It's such a godsend for writing OCs. Have some legally distinct Kaai Yuki.

>>26103 do you do local textgen?

>>26103 Based, I have to get back into textgen.

>>26104 I did, started with Mixtral finetunes (I bought a 3090 for local SDXL). They're okay, will probably blow your mind if you haven't played with LLMs at all. I've mostly moved on to using Claude because nothing else is even close for roleplay. Anthropic's mid-tier models (Sonnet, 2.1) are cheap and you can get free access if you know where to look.

>>26106 how touchy is claude about loli? i've got sonnet through scylla, just stayed away from anything underage out of concern that i'd be killing keys.

abmayo: https://mega.nz/folder/OMY2RRxQ#zYaokRZV361WSNImfrBCrQ/folder/nMAyTDII made with my new tag pruner code that uses the v3 taggers (on dev right now): https://github.com/derrian-distro/SD-Tag-Editor/tree/dev miku because it's abmayo, and the rest is hifumi as usual, can't seem to get away from her

>>26107 Absolutely fine, AWS Claude is hardly monitored if at all, keys usually live through the whole month.

>>26106 charlotte is cute what model/lora is this?

>>26109 neat

>>26110 forgor to use the extension... https://files.catbox.moe/7r9e3k.png https://files.catbox.moe/475z7k.png akai_sashimi from: https://mega.nz/folder/OoYWzR6L#psN69wnC2ljJ9OQS2FDHoQ/folder/W5AjzbLD antifreeze from this anon's mega: >>26108

>>26112 appreciated!

>>26106 >Anthropic's mid-tier models (Sonnet, 2.1) are cheap and you can get free access if you know where to look. Wasn't Sonnet 3.5 released for free? Is it worse than 4o and Claude? I honestly never did anything interesting with LLMs after I had a bit of initial fun with 3.5 when it released. I'm getting a bit bored with imagegen kinda, maybe textshit could be a good replacement for this. No local for me though, sadly.

>>26115 Sonnet 3.5 sucks dick and I can only imagine if they have a free option it's both heavily moderated and rate limited like 4o's

What could have been

>>26117 if only

>>26118 at least there are other characters worth making a lora of

>>26117 Lucy and Piper are great so i'm satisfied and apparently there's another loli coming next ZZZ patch too Just downloaded the lora for those two off civitai and was about to get to doing some proompting

>>26119 What a cutie, I will now play your game.

>>26121 cute concept

>>26122 sexo images and I still have some more images of that concept to touch up.

It's weird, but I think AI has actually made me less degenerate. When I first started genning, all I did was ryonafag shit. Now I'm increasingly just an /e/ fag. I'm also more concerned about composition and hand accuracy than about sex. Also, even though I burned out on like every other character, I never got tired of remiflan.

I think I spent half an hour autistically inpainting the teeth and finger here...

>>26125 meanwhile I haven't inpainted anything in a stupidly long time because all I gen are test images for lora I bake

>>26126 same but because i know once i start inpainting i'm going to keep noticing imperfections and inpainting more to the point that just genning more would have probably gotten me a better result sooner

>>26127 For me, it's more like obsessively inpainting minor details and shit only to notice after I'm done and posted it somewhere that some other way more important and obvious part is completely fucked.

>>26124 These are quality. It's clear that you care

>>26129 Thanks. The other thing I've been doing lately is revisiting prompts from when I just started with NAI and was fucking around to see if I can improve anything with what I've learned.

I've been messing around with foreshortening tag since I saw shuten anon using it and I was often wondering if it even did anything. Well, sometimes it does do something.

I don't even need the loss part of instant loss comics anymore. Maybe I never did. Maybe all I really wanted was pantyshots of cute girls holding weapons.

>>26124 Been pretty vanilla /e/fag for a while but recently started to gen more and more degen shit.

gosaitamanegi: https://mega.nz/folder/OMY2RRxQ#zYaokRZV361WSNImfrBCrQ/folder/bNoywaDJ

also abazi (not entirely sure if that is correct tbh): https://mega.nz/folder/OMY2RRxQ#zYaokRZV361WSNImfrBCrQ/folder/CMAmAbQI

Oh no no no no pencilsisters... https://github.com/lllyasviel/Paints-UNDO

>>26137 My favorite artist twitter melty was when some ai miku picture got popular and a bunch of artists made memes of miku saying pick up a pencil or something. The complete contradiction of using an AI character from an AI voice program (guess it's not replacing artists when they're just singers eh?) was totally lost on them, because they have no self awareness or thought at all. It was hilarious. Miku is everything they stand against from both a product and a character standpoint, but they never connected the dots because it wasn't their labor getting replaced.

>>26138 Kek, didn't think about that. I recently saw artists cancelling another one for supposedly tracing over ai gens, which is true because... well it just is ok?! Then someone posted "proof" that they are tracing by running their pics through i2i and people ate that up lmao.

>>26137 Real art died a long time ago. I hope the whole crab bucket of grifters devours itself.

>>26124 sexo >>26134 I should gen some Ro as well. >>26135 >>26136 Very cute, thanks

Apparently hdg janny didn't like a single post of mine but, the rest are somehow fine. For the taiga anon, a typo fix in the prompt. If they even see this. NAI still doesn't really like doing taiga gens if you don't include toradora! though.

>>26144 Probably deleted whatever got reported.

Wish i could edit the nai gens easily if only this shit leaked...

>>26144 She's saisaka now. I'm going to bully her forever!

>>26146 Inpainting with the metachs extension is pretty nice even on upscaled pics, infill usually manages to grasp context even on a cropped pic.

>>26144 The /h/ jannies are lazy jeets who just toss a 3-day on anything that happens to get reported. I got a 3-day ban for loli for THIS picture.

>>26149 funnily, i mostly catch a ban whenever i post characters other than my wife

>>26144 F, well I have been sent on a vacation for hebes.

Why is this greasy, smelly, socially awkward loser so cute?

>>26153 Can't believe she shaves.

>>26153 >last love messy faces like that and love Tomoko. She looks too clean though, Tomoko would never shave

>>26156 She might if she snagged a cool older boyfriend (who definitely isn't taking advantage of her), which will definitely finally make her popular (by demonstrating how cool she is).

This is the first time I've ever randomly gachaed into a red themed monochromatic image.

>>26158 it's a NAI error, sometimes it really fucks up the color when it's struggling to resolve a bunch of conflicting artists or compositions you can do some cool stuff with it

>>26160 cute wife

>>26161 my cute plap wife

>>26158 >>26159 I sometimes get really bizarre shit with NAI's SDE, actually shame I didn't save some of those.

>add lora trained on NAI gens to the mixture >local is now 10x better it's just sad at this point

>>26162 my cute showering wife

>>26162 nice. easy to inpaint her face into other girls too, sankyuu

>>26165 same but after >>26166 enjoy bro

>>26167 very cute

>>26167 This image is great

Good morning I hate NAI (the company not people using their product).

>>26170 good morning, i'm probably going to pay them $25 again soon

>>26170 >hi guys im a retarded leftist beggar everything should be free :))) anything else?

>>26172 Not him but they ARE a bunch of faggots for not releasing the papers on the training despite promising to do so. Also SaaS is ultimately dystopian and unreliable, can be taken away or break at any time and leaves you dependent on the service provider. I say this as someone who's subbed, but I can recognize why people take issue with services. Most people who dislike NAI wouldn't mind paying for getting the weights locally, but obviously this just won't work out.

>>26173 Weren't they going to release them when v4 comes out? Obviously they wont self-harm, if they do that and their userbase - thus funding - disappears we all go back to being eternally cucked. Whining about NAI is dumb, they aint even greedy money-wise. Just slow as shit, text behind and v4's still cooking

>>26174 >Weren't they going to release them when v4 comes out? Well, this is what they were claiming back in November-January, picrelated. Nothing has come out since then. There were also some claims from tabloida iirc but I can't find them right now. >Whining about NAI is dumb, they aint even greedy money-wise. Some people can't stomach anything non-foss/cloud and I can live with that, but the amount of seethe nai attracts is still a bit staggering, yeah. It seems to stem from the vg generals, they have a long history of nai wars.

>>26175 they spent like six months slowly dragging themselves to acknowledge that they can't compete with llama 3's 15 trillion token pretrain and they're gonna be stuck finetuning other companies' text models forever probably a bitter pill to swallow given how badly they got fucked by non-commercial licenses and how amped up they were to get a h100 cluster

Yes I like to gen fellatio.

It's just been a while since I genned this cute (not) war criminal

>>26178 >>26179 >>26180 sexo

It's over

>>26182 Counterpoint: It hasn't even yet begun

>>26184 Damn, that's one of the tightest Pony character LoRAs I've seen so far. And that cleft of venus is just *mwuah*.

>>26186 thanks! I try to put effort into my lora, I like for things to look mostly correct on my characters and styles, here's a link to the lora if you want to play with it, just a heads up, I don't use a negative or any score tags when I prompt, so if you do, your mileage may vary: https://mega.nz/folder/OMY2RRxQ#zYaokRZV361WSNImfrBCrQ/folder/GNhiFCwA as for that gen, it was a bit gacha, I kept trying to get her to keep her legs more together, as pony loves to spread them when it gets the chance. I actually was only able to get it to do closer to what I wanted once I started prompting for the shoes to be in frame, Pony is sometimes annoying to work with.

>>26185 We're so back and giwtwm >>26188 uooh

>>26193 You should tell her that without a job you won't be able to pay for all the anals it takes to gen her at this resolution. Pic 1 here and >>26191 are gorgeous btw.

>>26192 what do you use for these upscales in swarmui? i'm trying to get used to it and for the most part i like it more than auto/forge, but haven't figured out upscales do a degree i'm satisfied with yet. in your metadata i see "comfyrawworkflowinputdecimallatentupscalebynodescalebyds", does this mean you are not using the "generate" tab at all and are using the "comfy workflow" tab?

>>26195 I am, yeah, it's a bit jank because the swarm lora loader node actually breaks when you try to use the workflow in the generate tab without any lora loaded, but only then. I do also change the sampling to be AYS. It's a fairly simple upscale, 1.5x with a latent nearest exact at 0.6 denoise. It only uses built in nodes from comfy and swarm because I couldn't be bothered in the slightest to install custom nodes https://files.catbox.moe/x04bcy.json - workflow and an obligitory cunny post to keep on topic

plap

>>26199 very cute