/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: 15.00 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.

Uncommon Time Winter Stream

Interboard /christmas/ Event has Begun!
Come celebrate Christmas with us here


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

(2.18 MB 1024x1536 catbox_4hqhvd.png)

(1.98 MB 1024x1536 catbox_6bjgpa.png)

(2.04 MB 1024x1536 catbox_kumzle.png)

(2.07 MB 1024x1536 catbox_5e0qev.png)

>>14996 Thanks for pointing that out, I mix them up for some reason.
(544.14 KB 512x1024 catbox_jbh6vb.png)

(428.32 KB 512x1024 catbox_f5wm4v.png)

(615.73 KB 512x1024 catbox_x3adct.png)

(469.80 KB 512x1024 catbox_axpk2r.png)

>>14984 based, thanks
>>14984 way more work put into than I have but it is nice too see Good work anon!
I just realized it's been almost 2 full weeks since I've been able to just prompt. Guess that's what I get for prioritizing updates to my scripts over genning. Ah right I also have to look into fixing that bug with stealth_pnginfo and multidimensional as well.
>>15003 Multi-diffusion. auto correct screw you
(1.91 MB 1024x1536 catbox_1juxii.png)

(1.83 MB 1024x1536 catbox_pr1kl3.png)

(1.94 MB 1024x1536 catbox_2i2nmm.png)

POV: You knocked up the neighbors bratty daughter but she keeps coming back for more.
(1.98 MB 1024x1536 catbox_5rhjtj.png)

>>14825 managed to salvage the frip lora with dadaptation same mega https://mega.nz/folder/53gW3bbK#NxXyJ0l9T3JyAXolJfpBDw next is prolly shishi juuroku I NEED fat momijis
(1.06 MB 1024x1024 tmp0_zr_lyz.png)

>>14997 Oh wow, that shota lora is really good. I think my only gribe with it is that it tends to generate an older girl in pairings. https://files.catbox.moe/oqwjik.png
There are more retards than usual on /hdg/ now.
>>15008 Progress there seems to reset every few weeks. Still somehow miles ahead of most of the other boards, which is probably why all the tourist go there
>>15008 The containment, civitai keep getting down time after time. More even than usual it feels.
(1.52 MB 960x1344 catbox_cpxugg.png)

(1.43 MB 1152x1152 catbox_vgbbel.png)

(1.66 MB 1152x1152 catbox_5dll64.png)

(2.16 MB 960x1440 catbox_rubub5.png)

https://mega.nz/folder/OoYWzR6L#psN69wnC2ljJ9OQS2FDHoQ/folder/nl511QAb turnip kabura LoCON, I am not responsible for any cute and funny overdoses as a result of its misusage
>>15008 other boards too also n*musona tards keep posting their dogshit
>>15012 the other boards have always been shit, with /vtai/ possibly being the best intermediary location between sd/g/ and /h/dg while every other place is slow, dead, or uninspired.
(1.73 MB 1024x1536 catbox_dtuagc.png)

>>15011 will test seems promising
Is there a reason why the new Easyscripts isn't using as much VRAM as compared to the past? Did Kohya do something with his latest updates? I can train at higher resolutions now that gave me CUDA vram errors before
>>15013 got a 3day there for cunny a while ago sad I even censored with uoh stickers
>>15013 i brought it up last thread but /trash/'s /sdg/ is probably the best imagegen AI thread on 4ch. their interests naturally incline them towards being more creative with SD and there's surprisingly few trolls. main downsides are obviously you have to be tolerant of furry and there's not a lot of people doing training (in part due to there being no good finetune for furry/monster content).
>>15015 that is interesting. I'm not sure, I didn't even think to look at that when updating the scripts. Guess that's a good thing though.
>>15017 >no good finetune for furry content huh? there are
>>15019 furry exposed
>>15020 >not into fat canine knots NGMI
(1.70 MB 1024x1536 catbox_dz38ws.png)

(1.62 MB 1024x1536 catbox_y0v3t8.png)

(1.80 MB 1024x1536 catbox_dmy1v2.png)

(1.46 MB 1024x1536 catbox_mg7dri.png)

>>15011 >>15014 Locon approved I like it
(1.74 MB 960x1344 catbox_71yxkg.png)

(1.96 MB 960x1344 catbox_an7xoi.png)

(1.98 MB 960x1344 catbox_3friac.png)

(1.93 MB 960x1344 catbox_m8wcgo.png)

>>15022 nice anon, these are great for my next trick, here's two 2hu style LoCONs in one day, behold massakasama: https://mega.nz/folder/OoYWzR6L#psN69wnC2ljJ9OQS2FDHoQ/folder/awwlHBIC
>>15023 which Locon method have you been using?
(1.92 MB 960x1152 catbox_v0y5wi.png)

>>15024 locon_lycoris, 16/8 network 8/1 conv, d-adaptation works like a charm 90% of the time
>>15025 Ah I always kept conv dim the same as the main dim but is it recommended to divide it by half?
>>15026 That's what Lynn's notebook says so I went with it and it's seemingly been working well. Though Lynn's notebook says a lot of things of variable credibility so I should probably experiment with other rank ratios.
>>15023 HE'S ON FIRE
Are we going to have a language model thread here or are we just waiting for more uncucked models that are easy to run?
>>15027 I usually do dim16, conv dim8 for my style LoCon, You usually don't really need higher than conv dim8 for styles, and even dim16 conv8 you can probably do a fro resize and reduce file size further while keeping most if not all quality. That, I'd say, from my own personal experience in both testing locon and baking my own things, that it's a pretty credible, at least for that. though I don't know if it's worth it to train locon for characters, as I find they can have style bleed.
>>15029 probably should wait for things to be easier to run before we make threads for it, but in the future I'd personally be interested in it
>Gura has been streaming so less that her normie fans have been doing /vt/tier posts on reddit holy kek, this is funny I used to think the "GOORUH STREAM MORE" posts was only a 4chan thing
>>15032 >he visits r*ddit why'd you out yourself like this?
(1.74 MB 1024x1536 catbox_ysk13k.png)

(1.56 MB 1024x1536 catbox_glifxh.png)

(1.60 MB 1536x1024 catbox_xmsq2k.png)

(1.71 MB 1536x1024 catbox_pcsdtd.png)

>>15023 very cute nice work anon
>>15033 Reddit bait threads are posted on /vt/ all the time anon, I wouldn't actually post on that circlejerk hellhole that thinks wanting more from your oshi is as bad as rape
OK so did some testing, I can do batch 3 1088 with Lycoris LOCON baking, definitely a lot more compared to before so I think this update easy anon has done is the best so far, thanks a lot (but I am wondering what Kohya did to lower the VRAM usage)
>>15036 I'm happy that you like the update, but it's very likely something kohya did, not me. I don't exactly modify sd-scripts directly, just ensure that the correct values get sent to the scripts. But seriously speaking, I'm glad you enjoy the work I've put in, and hopefully next update has the UI I want to make.
>>15011 UOOOOOH CUTE AND FUNNY, CUTE AND FUNNY CUNNY ToT
>>15038 Looking forward to the UI you make, as long as if you keep the options you have with the popup and give us way more choices than the SD-Scripts GUI, I will be satisfied. Otherwise one note to keep, if you can have your next update automatically have the max bucket resolution size adjust itself when it's over the default 960 that would be nice. Not a lot of people train over 960 but just something to keep in mind
>>15041 great idea, but fun fact, at some point kohya changed the default to 256, 1024, so I plan on changing that in the scripts as well. I was thinking of going all out and setting it up to allow training of everything that kohya can train, each in it's own self contained gui of course. at that point I'd have to change the name from "easy_lora_training_scripts" to "easy_kohya_training_scripts" or something lel
off-topic but i love generating landscapes and i'm looking for a model that does photorealistic paintings, any suggestions? stuff like what michael james smith does
(248.54 KB 768x576 shota2.jpg)

(255.65 KB 512x768 shota3.jpg)

(341.87 KB 768x512 shota4.jpg)

(311.03 KB 512x768 shota5.jpg)

>>14997 this is great, gotta test if I can get it working with a character lora now or if it will turn the shota into a midget version of the character again
>>15044 anon-chama your catboxes...
(1.07 MB 768x1152 catbox_w7n14i.png)

(1.02 MB 768x1152 catbox_asb03i.png)

(1.09 MB 768x1152 catbox_rmx4qr.png)

(1.08 MB 768x1152 catbox_yy8dus.png)

I quite enjoying this turnip kabura lora
>>15046 mokouoooooh
>>15049 I'd take them if I were him
>>15013 The only reason the other places are slow are because people don't want to move. It's a thing that happens everywhere in 4chan where everyone complains the site is shit and has more people obsessed with shitting on it than people willing to discuss something but then you tell them to move to a less used board or a completely different site like 8ch and they act like that would be the end of the world.
Was gonna grab one from garbageai but it's down, anybody got a wedding dress lora?
Thinking about making a collection of resources for shopping onto an image before inpainting. Helps with adding dicks and cum, anything else you think would be useful?
Why are these hands coming out so good? I feel like AI is just sorcery where one day I have a bad case of hands then one day it's this
(1.37 MB 2688x4992 68061.jpg)

>>15058 It's already linked in that thread, but I found this site useful. You need to be able to read moon though. https://nijicollage.xyz/
>>15063 elf hayasaka?
(1.61 MB 3072x4608 68047.jpg)

>>15065 yes, indeed
(2.39 MB 1920x1280 catbox_al5tfz.png)

(2.09 MB 1280x1920 catbox_fpkrnc.png)

(2.31 MB 1280x1920 catbox_pieata.png)

Switched to figma lora first just because I figured it would be quicker. Again the whole dataset is being uploaded and extra trainings are there but I found 8-dim 8-alpha to be the most reliable. https://mega.nz/folder/2FlygZLC#ZsBLBv4Py3zLWHOejvF2EA/folder/mN9VzY6J Haven't done a ton of testing but the dataset is as clean as I could get it. Watermarks that I could find have been removed. Tagging was just run through the auto tagger so there's not really a reliable way to get the base and support to show up.
https://civitai.com/models/5891/crazy-expressions I've been liking this lora but it's not trained for NSFW and it really shits up gens sometimes due to it not being trained for it which sucks
>>15067 Nice, now generate more pics of the Aniki figma
(1.76 MB 1024x1536 catbox_5n6493.png)

(2.26 MB 1024x1536 00414-962994377.png)

(2.39 MB 1024x1536 00413-1833028759.png)

(2.00 MB 1024x1536 catbox_pkmmtg.png)

(1.99 MB 1024x1536 catbox_oal1c6.png)

Cute and funny overdose is the real condition
>>15073 Very cute, the topless + panties only combo is always gonna be the hottest shit but I do enjoy the camisole/bra + bottomless combo quite a bit as well.
(1.50 MB 1536x1024 catbox_9bj0dc.png)

what was the last thing your read anon? I mean you do educate yourself, right?
(1.91 MB 1024x1536 catbox_6z1guv.png)

>>15075 Bushcrafting, trap making.
>>15075 Omw to get educated
>>15078 >holding sword literally how? was it luck or did you use controlnet
>>15079 A solid mix of both. I could go more into detail, but basically: I got a lucky gen and then used controlnet and inpainting to keep iterating
(1.40 MB 1024x1280 00005-2129488584.png)

>>15079 Genning weapons without controlnet is a waste of time. You might eventually get a lucky roll, but you'd have to go through dozens of broken gens before that.
What's the difference between a lora and a lycon?
>>15081 I've seen your stuff on /g/
raymoo got some long toobs
>>15081 >>15085 I wouldn't mind a catbox for both of these, or link to lora if it's not on civitai
(1.46 MB 1024x1280 00043-428488615.png)

>>15084 Nice to hear that. /g/ is quite a shithole most of the time, sadly. >>15086 https://files.catbox.moe/rd24xd.png Pre-upscale version, upscale's got the same prompt, 0.45 animesharp
(2.41 MB 1024x1536 catbox_zfu45c.png)

>>15085 Ah another painting enjoyer
>>15082 lycon isn't an actual name, it's a mispelling of LoCon, which is LoRA that trains on every layer. Don't get this confused with LyCORIS, which is a blanket term that Kohaku uses to mean all of his models, which is LoCon, LoHa, ia3, Lokr, and his version of DyLoRA. to be entirely honest, I think Kohaku really shouldn't have created a blanket name, because that makes it harder to explain what is what, but to give you a basic understanding of what is what. LoRA = the original, you know this LoCon = LoRA that train on all layers, better for styles, might learn style on characters. if you can avoid that, probably would be great for characters as well LoHa = LoCon that are basically compressed, they work best when you want to put a lot of data all at one time, such as 16 outfits or the entire cast of an anime. ia3 = really small LoCon (300kb max I think), but don't really transfer between models well, trash, don't use them Lokr = really small LoHa, is at most like 2mb file size, but suffers from being weak in most cases, Only use if you really want to have a small model DyLoRA = short for Dynamic LoRA. I don't exactly understand them perfectly because I haven't read the paper well. but it seems like it's capable of producing better results for low dims, kohya's implementation of them aren't compatible with kohaku's extension so keep that in mind
>>15089 nta, but I appreciate the info
(2.29 MB 1024x1536 catbox_mg0gdl.png)

(2.27 MB 1024x1536 catbox_2043wr.png)

(2.15 MB 1024x1536 catbox_uoy04v.png)

(2.14 MB 1024x1536 catbox_dt38d0.png)

>>15090 No problem, lots of people don't actually know the different models because of the blanket term, so I take the time to explain when people ask.
(2.95 MB 1024x1536 catbox_pe4xuz.png)

>>15088 ya, I'm playing with the new shuffle controlnet, it's a lot of fun. >>15086 https://files.catbox.moe/yqxhm8.png You'll need starry night as the controlnet input image and this PR. https://github.com/Mikubill/sd-webui-controlnet/pull/785
>>15092 these van gogh reimus (especially the TUBULAR one) are so fucking good despite being so simple
>>15092 >>15085 does the new controlnet have that style transfer thing some chinks showed a few weeks ago?
(2.30 MB 896x1536 catbox_kg7l0v.png)

>>15093 I agree, have anotha >>15094 That one sucks because it's based on CLIP. The ControlNet author basically BTFO'd them. https://github.com/lllyasviel/ControlNet-v1-1-nightly#extremely-important
>>15094 >>15095 Here, look at this, left to right is no controlnet, T2I style controlnet (the one from a few weeks ago you're referring to), and shuffle controlnet from v1.1. Same settings for the two controlnets. It literally cannot compete.
(2.51 MB 1024x1536 catbox_yrtrr3.png)

(2.35 MB 1024x1536 catbox_qtlm16.png)

>>15095 Lovely
I'm desperate for a simejinameko style lora but I can't make it myself because of poorfag GPU
>>15096 If I crank the strength up to 1 for the T2I style adapter, it kinda works, picrel, but it doesn't really capture the style, and the only reason it can is because it's recognizing it based on CLIP. It basically sees the input image, can recognize it as Van Gogh, captions it as that, and tries to direct the output based on that. The method the ControlNet author used for shuffle is not using CLIP, and frankly I don't understand what voodoo he did but it works a shit ton better.
>>15096 shame, the chink one looked promising (weeks ago anyway)
>>15099 yeah i get it now (though i still like that result)
>>15100 Their examples were very misleading because they're styles CLIP can pick up on and they also had to use a second controlnet/adapter to maintain the composition. If you try plugging in some random artist's anime artwork it's not gonna know what to do at all because CLIP is terrible for that. With the shuffle controlnet you can throw literally anything at it.
(1.26 MB 1024x1280 00008-3688063988.png)

>>15098 Try colab if you have fake google accounts.
>>15098 >simejinameko shimejinameko is it, right? May look into it but if you provide a dataset it would be faster
>>15098 i'll make it for you if you give me a dataset (as many clean images with 1-2 characters in frame as possible) also how the fuck is there no jack LoRA
>>15102 >Their examples were very misleading because they're styles CLIP can pick up on and they also had to use a second controlnet/adapter to maintain the composition what a shame though it did look too good to be true
(952.77 KB 1536x1866 1681274678318114.png)

(59.56 KB 1920x714 1681275486151081.png)

I still have not had a chance to fuck around with controlnet, let alone the new stuff. Trying to read the dummyguide isn't helpful in learning some of the complex things i've been seeing like this.
>>15104 I'm not entirely sure how to make a dataset. Do you mean just grabbing as many images they have made?
>>15107 Just try using it yourself, I think it's pretty intuitive. Pretty hard to describe all use-cases and possible settings in a guide.
(1.70 MB 1024x1536 catbox_9slsra.png)

(1.71 MB 1024x1536 catbox_m83smq.png)

(1.80 MB 1024x1536 catbox_9nyq7d.png)

(1.75 MB 1024x1536 catbox_zjl3eb.png)

>>15108 it is like >>15105 said
>>15105 Alright I'm working on fetching the images right now.
>>15105 https://files.catbox.moe/pqqlxj.zip Shimejinameko dataset. I left out a lot of his earlier works since his art style really improved over the years and chose the stuff that isn't his sketches or rough linework
>>15112 https://anonfiles.com/l4k4L9leze/shimejinameko_zip mirror if the first link doesn't work
(1.99 MB 1024x1536 catbox_q4cdna.png)

(1.96 MB 1024x1536 catbox_jt2bq3.png)

>>15037 Style mixing sure is powerful.
>>15114 Wow you made that really fast
>>15116 still testing as I am not sure about this bake may put it back into the oven
(245.94 KB 1240x1754 20190502_5.jpg)

>Shimejinameko LORA in the works fucking based, I love the way that guy draws clothes
Speaking of based artists, was there a wagashi lora?
is shuffle gonna be powerful enough to replace style loras?
>>15121 oh wait nvm wrong guy kek
>>15120 Nah, it really only works best with heavily stylized images. It's even more fun to use it in combination with style LoRAs. You can't beat training for 1k+ steps with a one-shot solution like that. There is this which runs in <1s but I assume it's better at picking up concepts rather than styles. https://arxiv.org/abs/2302.13848
>>15091 >>15092 very pretty
>>15123 >Nah, it really only works best with heavily stylized images wouldn't styles like zankuro, csrb, dohna dohna etc count as heavily stylized? just wondering if this has potential or if it's really for those "i'm only gonna use this once or twice so i'm not gonna bother training" use cases
>>15119 catbox?
>>15125 Picrel is shuffle controlnet with dohna dohna artwork @ 0.85 strength, the onono imoko hypernetwork @ 0.85 strength (I've never found a LoRA for him that works well), both combined, and then combined with weights to my own taste. As you can see one-shot with shuffle isn't enough, but it works great in combination with their other artwork, or any other artwork you throw at it. So yeah, training still very much matters.
>>15129 ...aaand for reference this is without the controlnet and hypernetwork, so there's still a huge benefit to both to drive style.
>>15129 >>15130 so i'm guessing that for anime styles shuffle is more useful to strengthen a style lora or get a more unique style out of it and shuffle alone only works well on its own when you do extremely stylized styles like real painters'?
>>15131 Pretty much. I threw in some artwork I considered more generic and it can still vastly change results (https://danbooru.donmai.us/posts/6159895) But yes, definitely more useful when it's heavily stylized (https://twitter.com/SirayukiSIU/status/1636609589402832896) I put in a pic of a soyjak for the 3rd one for shits n giggles
>>15127 that's hand drawn art my man
>>15132 was hoping i could skip training some styles but oh well, no rest for wicked
(1.25 MB 1024x1536 catbox_7snym9.png)

subarashii
>>15135 now this guy gets it
(1.37 MB 824x1152 catbox_otqpt2.png)

(6.45 MB 2048x3072 catbox_lyra86.png)

>>15133 why even post it then
>>15139 the anon was asking if an artist lora existed, so posting an example of their art is a wise idea if it didn't since it would generate interest in doing so
>>15119 two or three of them even
>>15143 Thank you sir
(2.25 MB 1280x1920 catbox_0yq7yv.png)

Who's the girl next door living in the haunted mansion?
>>15092 is this using the style adapter in controlnet?
(447.31 KB 512x512 00105-2132915993.png)

>>15137 Cute Youmu!
Read what you replied to homie
>>15149 Meant for >>15147
>>15149 >>15150 yeah, I just read the thread. I didn't know there was a controlnet 1.1
>>15151 No worries, shit be movin fast
(329.46 KB 512x768 shota5.jpg)

>>15059 nice
>>15145 hand over the ashley lora and nobody gets turned into a spoon
(2.24 MB 1280x1920 catbox_81rz6m.png)

(2.49 MB 1280x1920 catbox_ies4v6.png)

(2.36 MB 1280x1920 catbox_q2l0b2.png)

(2.09 MB 1280x1920 catbox_rsie3i.png)

experimenting with fizrot + alkemanubis + ushiyama ame mixes added some csrb in so i can stop feeling like i should zero out, burn, smash then bury my ssd
>>15156 in order 0.3, 0.2, 0.3, 0.2
>>15156 >>15157 i like the styles i'm getting by messing around with all the loli-oriented artists but dear god they really ruin what little consistency the base model had when it comes to 2hu details on another note here's a regular b64v3 upscale with accidentally decent hands
(1.23 MB 1024x1536 00010-2707381757.png)

(1.28 MB 1024x1536 00011-2338137678.png)

(1.47 MB 1024x1536 hadashi-no-kenji-1030it-novae.png)

(1.85 MB 1024x1536 00039-2239416063.png)

(1.09 MB 1024x1536 00022-1439442639.png)

(1.41 MB 1024x1536 00010-2611109459.png)

(1.46 MB 1024x1536 00182-2707381752.png)

(1.66 MB 1024x1536 00003-4276655418.png)

(2.57 MB 1024x1536 catbox_4iosyk.png)

cute style, it's a shame the hands come out looking like ginger root
>>15162 didn't mean to spoil that, it's normal i promise
>>15156 2x 0.4 remacri, looks nice (just don't focus on the hands)
(1.30 MB 1024x1536 00004-2536168834.png)

Should I post civit-tier lora here if I am not planning on remaking it any time soon?
>>15165 if i wasn't okay with civit-tier loras i wouldn't have 60gb of loras
(1.33 MB 1024x1280 00010-452950379.png)

(1.21 MB 1024x1280 00047-1638515665.png)

>>15167 Forgot to mention that it's miyo_(ranthath) I think it kinda looks similar, but fucks up the eyes quite often, sadly.
(425.70 KB 576x704 added dick.jpg)

(553.46 KB 576x704 01680-2735792581.png)

(277.11 KB 768x512 2.jpg)

(448.91 KB 576x704 Illustration8.jpg)

I did it, so much easier to add dicks now https://mega.nz/folder/MrJ1WRLL#25hvOdtmUnhB-6miS8i-Cw
hands are still a mess but it's a nice style mercilessly cribbed the maid guy's preamble
>>15170 Imagine not getting a messy hand
>>15095 >Hires upscaler: Latent Diffusion upscaler? is this a different Latent Upscaler? for me it just shows "Latent".
>>15172 whens the last time you updated
>>15173 Ooh... Now that u said it anon I haven't for a good while after the gradio was fixed.
(1.85 MB 1024x1536 catbox_0ar9tr.png)

>>15170 based river prompter.
>>15172 It's from here: https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/4446#issuecomment-1312561298 https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/4446#issuecomment-1312598118 I implemented it slightly differently in my codebase because I think the way they did it was kinda dumb iirc. Honestly it isn't that great though, denoising kinda makes the difference less obvious between upscalers, and really I just forgot to change it to an GAN based upscaler, lol. Just use something else and you can get a very similar result. The only upscaler I've seen make a major difference and looks different than vanilla latent upscalers and GAN upscalers is LDSR, but that has a bug currently where using it will break the token counter. https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/9466
>>15176 Thank you anon. I'll take a look
i'm on hash a9fed7c364061ae6efb37f797b6b522cb3cf7aa2 should i pull?
(1.61 MB 1024x1536 catbox_y9f0lt.png)

(1.73 MB 1024x1536 catbox_vp08mx.png)

>>15179 Yes.
>>15171 Very pretty, you straight prompting for Iori in this or using a LoRA?
>>15117 slowly getting there moar testing will be done
(1.48 MB 1024x1536 catbox_c3gc04.png)

(1.62 MB 1024x1536 catbox_92a02q.png)

(1.76 MB 1024x1536 catbox_21i9bc.png)

(1.73 MB 1024x1536 catbox_7r4xvy.png)

>>15182 https://anonfiles.com/JaIeSflez7/shimejinamekoV5_1_safetensors Maybe some willing anon can help me test it so I can see if I need to rebake again?
>>15146 Catbox please? How did you achieve those perfect white pantyhose?
>>15183 nice, yeah I'll fiddle with it and give you my thoughts after I try it
(1.48 MB 1024x1536 00228-1290411949.png)

>>15183 https://files.catbox.moe/mnjnmk.png Did a Reimu test probably should have used other model, but whatever
(1.89 MB 1024x1536 catbox_blpdyz.png)

>>15187 weird my reimu always has her arm behind her head I wonder why :^)
toy posted about this but it looks like a solution looking for a problem tbh https://github.com/eideehi/sd-webui-better-prompt
I wanted to retrain all of my LoRAs using more recent settings but more than half of my datasets are unusable by my current standards and need to be remade from scratch. At least with later ones I learned my lesson but still very pain peko...
>>15190 >trained with a VAE all this time bros
>>15190 same story here, by now a few fanart works that I edited are gone from the internet because their artist was a menhera and took everyone down
>>15191 You're technically training with a vae anyways that's apart of NAI which is SD's basic VAE, the problem I have with all of it is that people are only doing it because of the "I'm too retarded to set a VAE myself"
(1.96 MB 1024x1536 catbox_ab8kmp.png)

(2.00 MB 1024x1536 catbox_17gkqy.png)

(2.00 MB 1024x1536 catbox_zvna1z.png)

(1.93 MB 1024x1536 catbox_qno7ss.png)

This is a pretty dangerous mix of styles
why am I getting a >A tensor with all NaNs was produced in Unet. Error with some LORAS? I have no-half-vae set
>>15195 i have a couple that give me that, my assumption is they were trained with a VAE that would cause you problems and the lora inherited that.
Basedmix anon, you still around? Haven't seen you post in a while.
>>15197 I've been posting recently but not much because I've been trying to understand language models and working on LORAs again so I can post outputs on the sites I use.
>>15198 Fair enough, so I take it the basedmixes are on hold for the time being?
(1.43 MB 1024x1536 00009-3186994817.png)

>>15200 as always I see the fucked up mouth after I smash "reply"
>>15199 Yes, it's been awhile since I last worked on based66 but I'll do it soon again, I'm wondering if the MBW page that one anon made which had real time updated info is still up. I should look into it also for block weights when it comes down to LORAs
>Nvidia fans coping hard with the 4070 >AMD GPUs in the 6000 series and above will soon have ROCm available for windows users soon
>>15203 someone on /g/ was getting assmad when they were told their 4070 sucked
>>15205 An anon is always getting assmad on /g/ but not as much as I've seen with /vt/ and bait regarding the new anime adaption for that series Oshi no Ko recently kek
>>15206 I have not checked in on anything /vt/ outside of /vtai/, whats up with them this time?
>>15207 Just bait regarding Oshi no Ko and the anime adaption, it was made by the dude that did Kaguya and not sure if anyone recalls the meltdown /a/ had about the manga but Oshi no Ko hits /vt/ the most because as a post there mentioned >Oshi no Ko's entire premise is that idol culture is built on lies and that parasocialism is bad. This is what has been already established in the west a long time ago, mind you (that's why I address JP fans). Thats why western idols/vtubers are so different from Japanese ones. Western vtubers being poisoned by irony and being unable to unironically tap into the parasocial idol culture is obviously a loss
>>15208 But anons just hate Aka as a mangaka and view him as one of the biggest hacks in the manga industry, most /vt/ anons are from /a/ anyways but the fact that this series has this premise just pisses off anons that hated him anyways
>>15208 I haven't read the end of Kaguya yet but boy do I remember the weird meltdown that happened when Prez and Kaguya finally fucked and the aftermath of the plot getting nonsensical. >idol culture bad Prez's dad being a streamer was funny, I had a feeling that the topic was gonna get expanded but not its own manga/anime kek
>>15203 i'm not coping, i have a 4090 :^)
>>15183 These look very nice
(491.21 KB 512x768 anna example.png)

(492.78 KB 512x768 aoi example.png)

(445.90 KB 512x768 rio example.png)

(529.81 KB 512x768 kotori example.png)

Trained a bunch of stuff over the weekend trying some new settings, think they came out decent. https://mega.nz/folder/pZ9nBJ5Y#cmpb-DWVSWqqXg3fi9zX2Q Part 1: Anna Kouzuki - Yugioh Zexal Aoi Zaizen - Yugioh Vrains Rio Kamishiro - Yugioh Zexal Kotori Mizuki - Yugioh Zexal
(463.76 KB 512x768 emma example.png)

(457.74 KB 512x768 mahiru example.png)

(500.99 KB 512x768 chisa example.png)

(472.75 KB 512x768 zexal uniform example.png)

>>15214 Part 2: Emma Bessho - Yugioh Vrains Mahiru Koizumi - Danganronpa Chisa Yukizome - Danganronpa Yugioh Zexal School Uniform Concept
>>15214 >>15215 Vrains had some good looking girls but such a nonsensical plot
>following illyiasviel's progress with controlnet 1.1 never stops being admirable to see such a high-powered guy have so much patience for such dumb questions
>>15214 >>15215 Those all look great, anon! Thanks. I love Rio.
(1.53 MB 1024x1536 catbox_9uh7j6.png)

(1.72 MB 1024x1536 catbox_kjhqpq.png)

(1.72 MB 1024x1536 catbox_ld6hxa.png)

>>15214 Pretty good anon, nice work
>>15220 can I ask where you got that model or how to make it, it looks really nice anon
>>15221 I was fucking around and just threw based64 and AOM2-7TthV3 A 40 pruned r34-SF1-1 together in the checkpoint merger, mostly because I used them alot
>>15222 nice, thanks
i think i figured out what bothers me about controlnet hands (at least the ones derived from the blender rig) they're too manly/nowhere near as feminine as they should be
>ywn pain
>>15225 wedding dresses were one of the first things I prompted back when NAI leaked, I should do that again
(1.01 MB 1024x1280 08206-91535808.png)

(1.17 MB 1024x1280 08188-2034691710.png)

>>15229 take responsibility anon
(2.01 MB 1024x1536 catbox_pg7d1f.png)

>>15230 Don't worry, I will.
>>15231 Now generate the classic family of five
>find artist I want to train >their art is ant sized and/or covered in SAMPLE Very inconsiderate to be quite honest family
>>15208 I haven't been on /a/ for about a year now but people are mad about oshi no ko? Is it just anime onlys being retarded again? I don't remember any seething in the manga threads when it first came out
>>15156 Don't suppose you got a fizrot and ushiyama lora download on hand, would you? I gotta get my 3 fairies fix.
>>15236 Never mind, imma dipshit and found links all by myself. To the fairy cave!
i shouldn't have bought a full tower what do i do a) get someone to help me lift it up on the table and potentially drop it on the way down b) further fuck up my back and build it on the floor
What're some current recommended GPUs to upgrade to if I'm using a potato of 6 VRAM atm? Would love to speed up my shit.
>>15239 fuck it i'm going to the hardware store to pick up some wood, i need to build a shelf for my DAS anyway
>>15241 you won the wings gacha in the last two huh
>>15240 What's your budget? If you can get a used 3060 12gb or find a new one at MSRP go for it, if you have a 1000 dollars to spend you might find a 2nd hand 3090 for under that
I can probably go up to 1.3 to 1.5k if you include shit like shipping costs or whatever.
>>15245 get a 3090
>>15243 today's my lucky day, the AI is even inventing new forms of lingerie for me
>>15245 >I can probably go up to 1.3 to 1.5k if you include shit like shipping costs or whatever. >>15246 Any particulars about the brand I gotta watch out for this kind of stuff? Or anyone goes for the 3090.
>>15248 Not him but DO NOT get Gigabyte EVGA is the holy grail.
>>15249 Well shit, my current potato IS a Gigabyte. Thanks for the warning.
>>15248 like >>15249 said, get an evga ftw3 ultra is your best bet, avoid the Ti if possible, not worth it even at used prices
>>15248 Avoid the 3090 ti not worth it, avoid Gigabyte try to get EVGA but if you can't the other brands are fine but do make sure that the person can provide the invoice and receipt for RMA, make sure to check the memory junction temps when you first get it while training because that's the most important thing you have to check out for with used cards aside from obvious broken cards from the get go
>>15252 and if you do get a mined card just hope the person you buy it from wasn't a retarded trend chaser and actually did the thermal mods that lowers temps, as long as if the person didn't get rid of the warranty covers you can still do an RMA
>>15253 Pretty funny I just found two EVGA 3090s for under 1k, man I went for the Ventus OC because they were the only ones up on offer when I was buying my 3090s, good luck anon 3090 is a card you'll want to keep until the 5090 and the PCIE gen 5 slots on modern motherboards will actually matter
>>15252 >>15253 kek mine has a year left on it but the seller lost the receipt, i'm fucked
>>15255 Pretty sure some companies will do the warranty from when the card was created, but knowing that 3090s aren't in production anymore yeah you're fucked if you don't have the receipt for used ones unless you go full retard and buy a new 3090 for the scalped prices
>>15256 which in my country one new 3090 ventus OC from MSI currently costs the price of two 4090s kek
>>15256 haven't been able to test it yet even tho i got it two months ago, if it's a dud i'll try to pester paypal or contact evga i guess
>>15253 Best fucking advice in terms of dealing with miner cards. >as long as if the person didn't get rid of the warranty covers you can still do an RMA I had to RMA a gigabyte 3080 TI after Elden Ring killed it, I had done the thermal mod on it months back after I saw how horrible the temps were awful and it still had issues. It didn't have any warranty seals on it when I did the mod so I sent it in, and got it back no problems. Funny thing is when I did get it/one back, the screws had warranty seals on it. Best part was is that the card STILL had thermal issues. I bit the bullet and just bought the MSI 4090 back in January and sold the 3080.
>>15259 I only plan on doing thermal mods myself after I send in my cards for RMA before their warranty is up next year, it's a good deal because I already know MSI is just going to send me a new 3090 that I can easily just replace the thermal pads and have those cards live up to the day when the 5090 comes out, not sure if I'll sell them because when it comes down to AI stuff VRAM matters the most anyways
>>15252 >>15251 Thanks for the help! Now I just gotta decide if I wanna risk the used route after all, newegg says every single used one they got is from China.
>>15225 >>15228 Whoa, these are great! How did you make them so detailed?
>>15261 For the used market it's better to just use ebay or your local pick up and buy used market, I went with the 2nd option because it was easier to haggle and make sure that I wasn't getting a shitty gpu
Pointillism LoRA is pretty interesting. You need to play around with the weight (>1 otherwise it might not even be noticeable) to get the most out of it. I can't get anything impressive if I use it with my overcooked LoRA but still, nice. One weird thing is that it cuts my IT/s by 1. https://civitai.com/models/26548/pointillism-painting-style
>>15262 Remacri at 0.4 did all the work, the base images are nowhere near as good It's a weird beast, 0.4 does wonders on most styles but completely ruins a few and you need to dial it all the way down to 0.1
(1.67 MB 1024x1536 catbox_x3d3nn.png)

(1.83 MB 1024x1536 catbox_6fk7r2.png)

(2.22 MB 1024x1536 catbox_q5vjv1.png)

(2.01 MB 1024x1536 catbox_5uc9us.png)

have I failed by making a lora that is unlike the target or succeeded by creating something new?
>>15266 I'll GLADLY take that, whatever it is
>>15267 https://anonfiles.com/feW7k5mfz4/mossacannibalisV2_safetensors I took all pics from https://gelbooru.com/index.php?page=post&s=list&tags=mossacannibalis and mixed them together but it seems that diverse range of styles was maybe a bit much
(1.73 MB 1024x1536 catbox_morxqi.png)

>>15268 but it is kinda nice even if the nose is more prominent than normal anime style
>>15268 I really like the end result even if it's not what it was meant to be
>>15264 I'm mixing it at weight 1 with CSRB at ~0.5 and holy shit that's nice. Complete opposite of what the style is meant to do but fuck, that's some REALLY nice skin and shading. Spoilers since they might be a bit too realistic for the board. Original vs Remacri 0.4, bless the upscaler for adding the extra string.
>>15265 Using hires fix? I barely get any new details when using anything other than latent, which tends to fuck something up, especially hands.
>>15272 Yeah, same steps as main prompt.
>>15273 Can't you just leave it at 0? It says it uses the same steps if you leave it alone.
>>15274 0.4 seems to change things for the better often enough and lower values are usually blurrier/noisier (duh) at <0.3 If you're talking about hires steps then idk, never tried leaving that at 0
>>15275 Yea im talking about hires steps
What are some artists that are good at drawing genitalia. Doesn't matter if they are well rounded or specialize at one in particular.
>>15278 my personal favorites are probably fishine, LEE, tamagoro, and nakagami takashi
>>15278 syuro :)
>sd/g/ is going full retardo again I wish this place was a little more active, I'm starting to hate every single sdg general on 4chins
(1.62 MB 960x1344 catbox_o0bgcn.png)

(1.45 MB 960x1344 catbox_5pxczy.png)

(1.46 MB 960x1344 catbox_i2zr8b.png)

(1.35 MB 960x1344 catbox_6ieom7.png)

remi is such a good test subject for loli style mixes
>>15283 What style?
>>15271 >>15277 >>15282 >>15285 catbox me on one of these homie, I too wanna try this LoRA mix
>>15278 aki99
miko armpits ToT
>>15289 Cheers, anon
(1.44 MB 960x1344 catbox_3phtbl.png)

(1.53 MB 960x1344 catbox_eof60w.png)

(1.47 MB 960x1344 catbox_6upren.png)

(1.31 MB 960x1344 catbox_yj6em7.png)

>>15284 https://gitgud.io/GautierElla/8chan_hdg_catbox_userscript/-/raw/master/out/catbox.user.js install this userscript and you can just right click on the image to see the prompts also, I think I have concocted a nutbladder-destroying style mix. it's over bros, there is no surviving this.
(1.79 MB 1024x1536 catbox_tqcvxb.png)

>>15290 Now I understand the reaction in >>15285, it really is a great feeling when you throw in several LoRAs at full strength and it Just Werks
>>15292 What's more impressive is that maybe two of them are from civitai. Can't remember where I got the disgust one.
>>15293 I assumed it was this one, glad I keep the original URL for all my random catbox LoRAs https://archiveofsins.com/h/thread/7213558/#7215154
>>15294 I have two in my folder and that's probably the one I autofilled in
>pointillism lora spits a key error >it's because a1111's ui doesn't support modern training methods >git pull >completely breaks wonderful. is there another UI i can try?
>>15296 It works fine on my end on a commit from last? month, check the extension Civitai page can't decide if it's a lora or a lycoris and it links to another extension but I don't use that one, this is what I have
>>15292 btw anon does that pointillism lora affect your IT/s WAY more than any other lora?
>>15298 Nope, runs at the same it/s for me. >>15296 It's LyCORIS, get the locon or lycoris extension. It worked fine for the anon in >>15297 because they're using the locon extension which supports the same <lora> syntax. I'm using the newer one (lycoris) so I have to use <lyco> instead.
>>15300 Odd, it cuts my IT/s by 1.something
Are tags like "looking away" or "looking to the side" or "looking down" almost completely broken? I swear they don't work except maybe looking away once in 10 rolls
>>15302 Or maybe I'm retarded and/or it's the combination of prompts I've been using but I've had trouble with the first two without any LoRAs loaded
>>15300 thanks, did git checkout a9eab23 to switch back to a working commit and that addon makes it work
>>15271 >>15277 completely uneducated speculation, but i've gotten some unexpectedly high quality results re: skin and lighting before from early step prompt editing changing large portions of the piece. i wonder if existing models are just overtrained on torsos and adding noise like a pointillism style breaks that up.
>perfect anchovy except the grin >upscale >don't really like the eyes but w/e >grin is even more fucked up >no grin lora PAIN
>>15306 just inpaint it
>>15307 I've been trying to avoid inpainting as much as possible. If I start inpainting over (more often than not) minor annoyances I'm never gonna stop.
>>15309 "it's almost perfect but" is the ideal reason to use inpainting
Why do we still not have a decent inpaint guide?
>>15310 I know but consider the following: I know I can win the gacha but I also know that even if I do win it's impossible for the result to be perfect - which means I can stop spinning the wheel whenever I get something I like *enough*. I don't think I'd have a clearly defined point of diminishing returns with inpainting and I already know I'm just gonna spend WAY more time on it vs just spinning the wheel and calling it a day. tl;dr autism
So I'm wondering, does cosine with restarts have to correspond with your epochs? I noticed someone saying a theory that might be the reason why someone's LORA is better at eg; epoch 7 rather than epoch 6
How the fuck do you use hiresfix from the img2img tab?
(1.11 MB 1144x800 00133-2034500318.png)

>>15311 hit me up if you find one
>>15315 >xray is that a lora?
>>15316 I know of one, but I haven't read all of it so I'm not entirely sure if it works well. https://imgur.com/a/3FDYMTk
>>15319 >Dec 30th oh boy, this can be either scrolls of lost ancient knowledge tier or lol what the fuck were we thinking back then tier. I'll give it a read after im done stuffing my face.
(5.19 MB 1536x2304 catbox_984fh4.png)

spent all day trying to find a good lora blend for masamune shirow and came up empty but at least this one is kind of cool
>>15320 (me) Yea its a decent guide, it at least explains what each knob and slider is for and when to use it for both standard Img2Img and inpainting. Not sure how much is outdated and what stays the same
>>15288 >(medium breasts|flat chest) This is supposed to be using [] instead
(1.59 MB 1024x1536 00044-1645599276.png)

(1.81 MB 1024x1536 00183-1063486680.png)

(1.71 MB 1024x1536 00017-2795406663.png)

(1.74 MB 1024x1536 00010-393488160.png)

(1.53 MB 1024x1536 00018-2692289208.png)

(1.66 MB 1024x1536 00014-39811765.png)

(1.73 MB 1024x1536 00457-2588399268.png)

(1.82 MB 1024x1536 00026-3549671467.png)

(407.64 KB 512x768 catbox_srjzij.png)

(434.14 KB 512x768 catbox_jkr763.png)

(514.97 KB 512x768 catbox_o43rfl.png)

(492.87 KB 512x768 catbox_1z30y6.png)

"A new kemono friend appears"
(904.07 KB 960x640 catbox_6135eo.png)

(802.25 KB 960x640 catbox_bmeop1.png)

(803.37 KB 640x960 catbox_fsk7m4.png)

(774.30 KB 960x640 catbox_xw37nj.png)

Turns out the hardest part of making a lora from anime screenshots is figuring out how to use a huge dataset. My solution is sort similar images info folders and equalize repeats using the least common multiple. Results in a huge number of steps for one epoch but that doesn't matter if you're only doing a set number of steps. It seems to work but need to test more.
>>15331 Oh hey I loved this anime's style
>My solution is sort similar images info folders and equalize repeats using the least common multiple lmao what, prune similar images and crop the remaining ones using ML solutions. that's literally all you need to do lol
>>15268 IDK if you care, but there are a lot of images missing from gelbooru that can be found at https://imhentai.xxx/gallery/574456/ https://imhentai.xxx/gallery/654981/ Though some are just sketches especially early in the first collection
>post deleted on /vt/ >can't even see why since router restarted since Jannies stay loosing, as do chubby haters
(1.79 MB 1024x1536 00034-3642377046.png)

(1.83 MB 1024x1536 catbox_funhx7.png)

(1.87 MB 1024x1536 catbox_7imh4f.png)

(2.01 MB 1024x1536 catbox_kn8eh7.png)

(1.52 MB 1024x1536 catbox_ggn7ga.png)

does anyone have a good guide for tagging data? I want to put more effort into my LORAs
>>15339 There was someone who made their workflow in a previous thread. Biggest factor is you will need to manually tag your stuff.
>>15339 https://rentry.org/lora_train#captionstags I have a general guide on how I tag things, long story short, I generally use BA Anon's style of tagging
>>15339 smilingwolf personally recommends the swinv2 model for danbooru tagging things, if you're tagging stuff that's visually noisy or has elements you think might be mistagged make sure to check a handful of the generated caption files DO NOT FUCKING PRUNE YOUR CAPTIONS, NAI WAS TRAINED ON DANBOORU TAGS AND IT PREVENTS THE AI FROM MAKING ASSOCIATIONS WITH TOKENS ALREADY IN THE BASE MODEL. that being said, adding new tokens to define different outfits is fine, just DO NOT PRUNE SOMETHING IF IT IS IN THE IMAGE.
>>15338 more friends are always good... now we need crested ibis so I can make them be gay
>>15342 I would only suggest getting rid of redundancy and false positive bad tags, but don't do the bullshit that retards in /h/ have been peddling like removing hair eye color and shit like that.
>>15342 >>15344 >without pruning : hair style and eyes don't get genned correctly >with pruning : they do simple as
>>15345 sounds like a skill issue because none of my character LoRAs are pruned and they work fine
>>15346 because you add hair and eye color in prompt
(583.97 KB 640x960 catbox_m8ecr2.png)

(670.89 KB 640x960 catbox_eoxctb.png)

Quick sirfy lora https://mega.nz/folder/2FlygZLC#ZsBLBv4Py3zLWHOejvF2EA/folder/WY9ywK7Q Result of an experiment to see if you could use cache latents to disk as a way to share datasets without the original images. Unfortunately looks like this isn't something that's possible at the moment.
>>15337 actually better then csr own art, his style is so ugly but the ai fixes it
(3.30 MB 1536x2304 catbox_y86m9d.png)

did you find something down there, master?
>>15351 I HATE THE LYING ON COUCH PROMPT I HATE THE LYING ON COUCH PROMPT I HATE THE LYING ON COUCH PROMPT I HATE THE LYING ON COUCH PROMPT
>>15349 yeah, i recall csrb artwork being kinda mushy but the AI gives it better form
>>15353 imagegen's obsession with faces does the style a lot of good too, instead of csrb's usual 90% of the picture is back or abs and maybe there's half an eye in there somewhere bullshit
(761.27 KB 1792x3328 68065.jpg)

(864.29 KB 1792x3328 68064.jpg)

(1.31 MB 2688x4992 68070.jpg)

(1.81 MB 2688x4992 68063.jpg)

>>15355 best* girl, hand over both loras *depends on the day, sometimes it's kei
>spent four hours playing with lora mixes before i noticed some faggot's picture i imported applied a clip skip 1 override
I hate the override settings
>>15360 being visible doesn't make it any less annoying that I have to disable overrides every time I load an image from pnginfo
>>15362 is it really an issue? most people here used b64v3 and the clip skip override sounds extremely rare
>>15363 the rareness is why i didn't notice it or think about it for four fucking hours
>>15364 sorry anon but
>>15360 This previously was an inssue for me until I changed the order of this to be fixed to the top in settings because previously it got shoved below all extension accordions off screen. Now I don't mind it.
>>15366 Excuse my typo, have a gen.
>>15360 So where in my fucking post did I say I couldn't see it? Make that stupid box and arrows on my fucking post nigger.
>>15368 oh it's this twitchy sperg you really can't handle the slightest tease huh?
>>15369 >this twitchy sperg who? >obnoxious image >"slight tease" sure buddy
(2.06 MB 1024x1536 catbox_w07xaw.png)

>>15367 at least i have a bunch of potential styles to mindlessly churn through for a while
(240.95 KB 960x640 catbox_32l035.png)

Been doing voice gen stuff and of course the only decent models available are pony. At least they've got John de Lancie. Anyway have Lagtrain while I'm training a Nanahira model
(1.76 MB 1024x1536 00100-869122906.png)

(1.93 MB 1024x1536 00125-1407523815.png)

(1.86 MB 1024x1536 00095-2283686476.png)

(2.01 MB 1024x1536 00008-2144409303.png)

(1.85 MB 1024x1536 00119-3293090893.png)

(1.88 MB 1024x1536 00100-4007193288.png)

(1.90 MB 1024x1536 00020-2394469013.png)

(2.02 MB 1024x1536 00019-4054142530.png)

>>15376 i LOVE the last one
>>15373 what's the status on good voice gen? i don't want to give 11labs my money but at the same time free shit sucks and 15ai is almost literally a fucking joke
>>15378 Depends what you want. I haven't really looked into tts but for me I just want inferencing on existing voice recordings so so-vits svc, diff-svc/fish-svc are the go-tos. Only recently has realtime inferencing become available for so-vits with so-vits-svc-fork or RVC so that's my current interest. Here's bugs bunny voice lines inferenced with a Pekora model pitched up half an octave in-program with sovits fork. https://voca.ro/12WD6ufVLcGV
please don't spook the birds
TINY ANGRY BIRDWIFE EROTIC
heun vs DDIM for taking quick samples before upping step count? why?
>>15382 isn't heun slow as shit? why not euler?
>>15383 euler is non-deterministic. heun is slower yes, but it also arrives at convergence much sooner so you can get away with a lower step count from what i've read
>>15357 Choose your words carefully, big guy. This is my wife you're talking about here
>>15385 >big guy For you.
>>15384 >euler is non-deterministic Simple euler is deterministic though? euler a isn't
(2.00 MB 1024x1536 catbox_rmeh6a.png)

tiny happy blindwife erotic
uuuuuoooooohhh my fucking DICK I'll rebake this lora without flip aug though I don't think I'm going to get that fang to come out without a fang lora but god damn MY DICK anne from the new pokemon anime btw, made entirely from screenshots
I don't suppose there's a Yaia from Granblue Fantasy LoRA floating around out there? New unit means new art, so wanted to ask if I had to try and make my own.
(2.03 MB 1024x1536 catbox_dir0vl.png)

(1.95 MB 1024x1536 catbox_q3v9z9.png)

Cute butts.
>>15342 I autistically tagged every detail in my dataset and prompting custom outfit words does fuck all unless I prompt every single detail of the outfit, and even then, parts of other outfits tend to leak in when they share tags. Though the fact I didn't try to balance them out with repeats might be contributing to this, so out of like 8 main outfits and a bunch of random ones without a custom keyword, the outfit that has the most images is like 33 out of 150. But even then, when I really use all the tags, the outfit comes out mostly accurate, with some smaller artifacts here and there, like doubled bows and ribbons or jingle bell on the chocker coming out more like cow bell. It's pretty versatile though. I might try pruning the tags a bit next time to see if it's any better, but I'm too lazy and I'm still pretty satisfied with my first lora, I like to generate unofficial outfits more anyway. Just some experience and thoughts from a training noob, if anyone cares.
(1.64 MB 1024x1536 catbox_a4pvx1.png)

>>15388 Ah, I see you're a man of culture as well
>>15388 Link to lora?
Okay rebaked anne lora It's limited but about the best I can do right now until there's more art and/or appearances. It makes my wenis penis happy so I'll share it any way https://mega.nz/folder/dfBQDDBJ#3RLMrU3gZmO6uj167o-YZg/folder/kDByBZoA
Editing tags for better LORAs is annoying, I can't take it anons, it's literally the last thing I need to put effort into for better LORAs
>>15398 Now imagine a finetuner and 200k images
>>15399 Yeah I can't imagine that, the best I could do is try to edit the common tags into more specific tags but the "tagged once out of a 100+ images" shit is the stuff that gets me the most because you wonder if it's better to keep it and try to make it a specific common tag for the rest of the images or not, editing images and upscaling was the easier part of training for me but who knew the tagging part was going to be the biggest obstacle for me
(736.78 KB 3840x2160 00030-1451871461.jpg)

(688.37 KB 3840x2160 00031-1451871460.jpg)

(706.28 KB 3840x2160 00020-2817811012.jpg)

(798.47 KB 3840x2160 00032-1451290871.jpg)

rivers
*out of memory's ur CUDA* hehe
(2.02 MB 1408x1664 catbox_uu9p4k.png)

>>15394 Every Toph, Everywhere, all at once.
>>15407 >publicly impregnating my wife in front of my wife, my wife, and my wife what a nice dream
Is there some secret to generating good images in landscape? SD seems to dislike doing characters in anything but portrait
(659.39 KB 3840x2160 00034-1484758859.jpg)

(664.87 KB 3840x2160 00049-2315109838.jpg)

(686.11 KB 3840x2160 00033-2106943475.jpg)

(617.57 KB 3840x2160 00023-1451290878.jpg)

lakes are nice too
(2.12 MB 1280x1920 catbox_ab6jfm.png)

This lora is going to torment me, I didn't have this much trouble getting basic character details with the hyouka lora. At least the style is good
(2.04 MB 1536x1536 catbox_h5cudz.png)

>>15408 Toph is worried that this might impact her career trajectory. Ah well, how hard can raising kids be? They'll probably be fine if she just pays absolutely no attention to them whatsoever - after all, that was how she was raised and she turned out great. There is no possible way that this could go wrong.
Average day on the beach with Chitose >>15410 >>15401 Nice water prompts. Gotta appreciate the classics
>>15412 are you talking about korra? i watched one episode and forgot about it.
(4.46 MB 1920x2816 grid-0062.png)

(4.61 MB 1920x2816 grid-0061.png)

(9.73 KB 566x820 test1.png)

if anyone remembers, like four or five months ago there was a short-lived thing whereby feeding a simple geometric sketch into img2img at very high denoise, like .98, would consistently generate more interesting poses and composition. this died quickly because img2img didn't have hires fix so it was kind of assy to work with. attached are a few surviving examples, i deleted the originals and the img2img input but it was just a couple of black triangles on a white background like so.
>>15409 guess not
so apparently xformers 0.18 is causing unet nan errors in the webui
>>15417 Theres one on civitai too I think but this one's probably better knowing that site
>>15425 In the end I decided to use the ones on MEGA. 64 is a bit duller and the eyes are a bit too on point but it gets the headband right more often. Had to tone both down below at least 0.8 since they were interfering with the seed and style mix.
>>15409 Latent couple, regional prompter, multidiffusion, paint-with-words, GLIGEN -- there's a lot of options.
>>15427 I meant more like.. tips and such for doing it without extensions/without needing to do prep work like with paint-with-words. Maybe some common sense stuff I'm missing. My problem with it is that I get blurrier, less detailed, way more "smudged" gens, etc. Plus the chance of getting wonkier anatomy/ghost limbs more often. Depends on the model but the results are usually similar.
>>15428 >without extensions I was going to say img2img but then you mention >without needing to do prep work There's really no good way otherwise and you just have to get lucky. It's the equivalent of trying to do zero-shot prompting with LLMs. Unless the model is stupidly powerful you usually need to guide it from the beginning somehow. Latent couple is even referred to as two-shot diffusion.
>>15429 I've seen quite a few good landscape gens on here during the last few months, maybe there's something I'm missing? Do you need to change your sampler/steps or have a certain resolution? I'm trying 768x512, maybe that's not optimal? Do you need to change your prompting style (tag order, doing blocks with BREAK, etc) and/or use certain tags, etc. Stuff like that. So far I haven't really changed any of that, I just do what I do with portrait res and switch the resolution around.
(141.43 KB 1080x590 1682067131675499.png)

Using SD on jewgle colab is now against the ToS
>>15431 They're good, just wish the guy would rebake them with the new tech so that we don't have to choose between flexible with good colors vs accurate with worse colors
>prompt ran >"hmm medium breasts isn't cutting it" >"big breasts" isn't a tag according to deepbooru >it's "huge breasts" instead >"eh fine, how big can huge be lol" PRETTY FUCKING BIG THAT'S HOW BIG
>>15435 >"big breasts" isn't a tag according to deepbooru The real tag is "large breasts".
>>15430 They're most likely using one of the methods I previously mentioned. Posting one of my own to run my point home further.
>>15437 >takorin HAND THAT SHIT OVER RIGHT NOW
>>15438 Not mine, someone from /h/ shared about a month ago. Definitely one of my favorites. https://mega.nz/folder/JhlQwIxR#sunbu0qeSPhP2REqylIhxg
(1.25 MB 498x421 aqua-aaaaa.gif)

>>15439 >A MONTH AGO WHY WASN'T I TOLD AAAAAAAAAAAAAAAAAAAAAAAAA
>>15437 hand the catboxes over too pretty please
>>15439 >>15440 I'm GENUINELY upset, takorin makes my dick go thermonuclear and I asked if a style thing would be possible every time a new major thing popped up (lora, locon, loha, lycoris, etc) and got "no it's not worth it, it's not gonna look like his style/it'll look too smooth/you'll need to fiddle with pixelization scripts all day" etc every damn time.
(1.19 MB 1024x1536 catbox_hhyked.png)

(1.29 MB 864x1536 catbox_avdvme.png)

(940.33 KB 864x1536 catbox_422vmi.png)

(207.08 KB 512x512 catbox_e3nvxg.png)

>>15441 Sure, have a few additional ones https://litter.catbox.moe/bd4cxf.png https://litter.catbox.moe/3i7q1g.png >>15442 My bad, anon. I was even in the thread on here when someone mentioned making a takorin LoRA but I nor anybody else seemed to take a shot at it, so my eternal gratitude towards whoever on /h/ did. I don't think they're here.
>>15443 I think I'm the one who mentioned wanting to make one. >it works it's over (for my dick) >we can gen takorin without niggers WE'RE SO BACK
>>15444 If I load stuff from pnginfo and my default hires steps value doesn't get overwritten does hat mean the hires steps are set to 0? Can't find it in the image's metadata
>>15445 Yeah alright, 0 > matching prompt steps in this case (tested it exactly once so it's a fact now)
tfw too stupid for regional prompting, no oyakodon and sister x sister gens for me.
(1.30 MB 1024x1536 00027-2618410951.png)

(1.48 MB 1040x1536 00006-319904734.png)

(1.25 MB 1024x1536 00107-707572566 (1).png)

>>15440 >>15442 crtl + f in the gayshit/put your favorite artist's name in archived.moe right now anon. You'd be surprised how many loras are stealth dropped then buried by other posts. That said, I was lucky enough to be there when the guy who made it shared lol
>>15432 Seems like google is starting to crack down on colab but here's a few takeaways: - Google is not banning users (Not yet at least) - They're mainly upset about the remote connections and not stable diffusion itself - PRO users can still use colab without restrictions or warnings - PRO users without compute resources "will not be banned but politely dissuaded from continuing" - This is only part 1 of their measures against SD users
>>15375 >>15449 >They're mainly upset about the remote connections and not stable diffusion itself no, they're mainly upset about the out of control gpu usage, which let's be honest, is the work of sd. i'm surprised it stayed free for as long as it did and in google's position i would've done the same thing
>>15449 Damn poorfag bros can't catch a break
(1.72 MB 960x1344 catbox_mpudkn.png)

>>15439 what the fuck how did I miss this one, good shit
any anons running Vlad Diffusion? It has Lycoris support and apparently training too
Have people unlocked the forbidden fruit of generating specific melee weapons/firearms well yet?
what's wrong with my syntax here? this is from inside a text file using the built-in a1111 "prompts from file" script. i've copied syntax from various github issues, but my SD refuses to gen anything now matter how i set up the text file: --prompt high quality, anime screencap, anime colors, stuff3 --prompt high quality, anime screencap, anime colors, stuff2 --prompt high quality, anime screencap, anime colors, stuff1
>>15456 I'm not exactly sure what you are trying to convey.
>>15457 inside the automatic1111 UI, under the scripts section, there's a script called "prompts from file or textbox". the lines i posted are located inside a file, and no matter how i've changed the syntax inside the file, the prompts don't get uploaded from the file into automatic1111 when I upload the file into a1111
>>15454 I have a genuine hatred for the entirety of that fork. all of the defaults are dog shit, the theme is god awful, and it forces you to install a bunch of completely useless extensions, and no you can't remove them, they literally are built into the core of the code. I personally think it's not worth even trying to use, because I ended up having to spend a good 40 minutes changing literally everything about the UI, from design to settings, just to make using it bearable.
>>15459 yep, copied it verbatim and SD still aborts the generation at 0.01seconds in. i'm using firefox on linux so maybe that's the reason? refuse to use any chrome shit though
>>15448 I stopped visiting /hdg/ daily at around #235 and the last one I looked at was #258, that should tell you how many potentially great LoRAs I'm missing out on. I just assumed someone else was keeping an eye on both threads and that they'd post the good shit here too lol
>>15460 oh, and it definitely doesn't have lora and lycoris training
>>15436 I mainly use (((sankaku))) so big_breasts is probably an alias for huge_breasts
>>15463 huh ok, going to ignore it then thanks.
>>15454 I also have a deep hatred for that fork like >>15460, though it's mostly because the author is retarded. He already does a lot of pull requests on the AUTO repo, and he's been the source of a LOT of bugs. Shit like making grid unusable because he NEEDED to put decompression bomb prevention, but removed (not just forgot) the factor to convert megabytes to bytes. He literally doesn't test his PRs
>>15466 i checked the guy's twitter and linkedin, looks like some middle aged "big tech" codelet who just wants some clout and recognition
is there a way to disable controlnet for the hires fix pass? i've been playing with using shuffle at high weight and short steps to strongly color an image, but it applies the shuffle map again at the start of highres fix and that tends to totally fuck the anatomy.
>>15466 >>15467 so it's just a clout chaser? That's lame, I was hoping for more UIs aside from Comfy just in case Automatic111 gets nuked somehow
>>15471 will never get nuked + it has god knows how many forks and you can always just archive a fresh install of a good commit comfy has a ton of potential but everyone's too busy with the gradio garbage cause muh sunken cost (other extensions) and muh learning curve (even though you can learn just by importing webui gens) i tried invoke, it's an absolute fucking slog to install even through the automated installer but it's cool, workflow seems totally incompatible with our coomer workflow though, think it's more for post-nut clarity art
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
>>15472 I would absolutely love a non webui system that is just as easy to use... actually, when I was looking around for stable diffusion webui replacements I found this https://github.com/arenasys/qDiffusion seems pretty decent, UI seems to not be really annoying, doesn't seem complete though
>>15475 I was about to say "i'd even take some shitty hacked together qt ui over gradio, i doubt it could be as bad" before clicking the link and yeah that's exactly what I meant
got some experimenting to do, see you next week (probably in less than one hour)
>>15476 to be fair, it seems pretty early in development. if it gets well known enough I'm sure people would be willing to help out.
>>15479 >>15480 huh, curious override setting for the model seems to stop working at random i thought the first upscale looked weird since it was supposed to look more like the second one so i looked at the console and yeah, the model override setting started working again when upscaling but it wasn't working when i generated the original pic
(3.18 MB 1280x1920 catbox_1h00iy.png)

(2.87 MB 1920x1280 catbox_a6zp3p.png)

(2.67 MB 1280x1920 catbox_usuz84.png)

(2.64 MB 1280x1920 catbox_m28bxh.png)

Another quick lora in between training Nanahira sovits and still processing nendoroid images (still) Puchiya (Hoshino Fuuta) https://mega.nz/folder/2FlygZLC#ZsBLBv4Py3zLWHOejvF2EA/folder/uMsUwK6K
>>15044 >>15007 the mega link is dead now, can you reupload?
some fucking nigger on 4chan got tired of spoonfeeding and tried to be both funny and passive aggressive at once by linking here it's over
>>15484 gayshit has always linked here
>>15483 https://mega.nz/folder/8uw2jQoB#-nv01mKy0dxDpQnyWYEhNQ I didn't download all of them just these two
>>15485 yeah but 4chan niggers are retards who can't read to save their lives and most people just ctrl+f instead
>>15475 I'm working on an alt UI to ComfyBox but I'm in the middle of nowhere without a good computer so it'll be another week before I can work on it again. So basically my idea is to take ComfyBox's default frontend and slap a presentation mode on it so you can hide the graph away, I always found interacting with the graph really annoying and tons of visual programming systems from decades past have already solved this usability problem. My goal is to have a replicable version of voldy's UI with tabs for each workflow and extensions of some kind. I'm still mulling about how to improve the mobile UX also, because using webui with a bad connection is almost impossible since it locks the generate button half the time. I'm thinking instead of going with a native framework I'll just reuse the same web-based UI because that way extensions can have one shared codebase. But I want the usability to be native-like for my own sanity even if the UI won't be native. If this is going to be "the" UI for ComfyBox/SD then I wonder what other people will want out of it so I don't waste my time designing myself into a corner. So any thoughts? Especially regarding how the UI builder will work. For simplicity I organize the panels into three vertical columns for now, but I've sometimes wanted horizontal alignment also like in a grid.
dead internet theory but it's bots making loras and uploading them on civitai instead
>>15468 >disable controlnet for the hires fix pass gen a lowres pic and img2img upscale
>>15489 Fuck I meant ComfyUI, yeah ComfyBox is the name of the project right now, I doubt a lot of people will get confused, but who kniws
>>15492 the results were uninspiring even when i got a little weird with it like genning a good small picture, sd upscaling it, then using it as a controlnet to regen the same picture at high base resolutions that would normally break. latent upscaling is just too good.
>>15474 >>15479 It's fun to imagine what the animations would look like
>>15495 I don't think we'll be able to generate 1:1 animations yet but with all the controlnet stuff I've seen on toyxyz's twitter I think we can get *close enough* right now
>>15496 I don't think we'll be able to generate 1:1 animations soon* I hate it when I change my mind mid-sentence and forget to correct the rest
>>15494 Well just use latent upscale in img2img?
>>15496 Oh yeah of course, just the thought is fun. Animations are a bit far off still since high quality tagged datasets are scarce for such a thing, even NVIDIA is just using Shutterstock videos currently.
>>15499 Didn't mean through text2video, meant through autistic controlnet pose fuckery like this (and with the appropriate loras/extensions) https://twitter.com/toyxyz3/status/1641599874448891905 Just need to find a way to improve the outputs' consistency but apparently that's already being worked on? https://twitter.com/toyxyz3/status/1649230593337757697
>>15498 doesn't seem to work the same way
>>15502 >latent upscale 0.45 denoise don't use latent below 0.5 denoise >doesn't seem to work the same way Add "upscaler_for_img2img" to your quicksettings Check "With img2img, do exactly the amount of steps the slider specifies (normally you'd do less with less denoising)." setting >hires fix 0.45 denoise remacri Use "just resize" option after you apply settings I mentioned earlier Use same seed and see if now it's the same (xformers shit aside)
>>15468 What do your controlnet parameters look like exactly? I've been changing the starting control step value to 0.2 and leaving the end at 1.0 and that has worked out great for me even with hires fix, since then it mostly keeps the composition you would get without the shuffle controlnet. If I do the opposite like I assume you mean, putting the strength really high and changing the ending control step to 0.2, that seems to mostly produce garbage anatomy for me even without hires fix (the image you posted is pretty subjective to this too).
>>15504 Oh you did mention using latent upscale in hires fix, that's honestly where most of the issue is gonna come from I think. I did a bit more testing and even if I do get a good gen using the latter settings, hires fix won't mess it up for me, but I don't use latent upscaling.
>>15505 yeah latent is very hit or miss
>>15468 and my whining has been answered, "Use only-mid-control on high-res. fix (second pass)" in controlnet settings knocks off about 90% of the controlnet's effect on the hires fix pass, looks like. >>15503 hires fix remacri at 0.6, hires fix latent bicubic at 0.6, img2img just resize remacri at 0.6, img2img latent upscale remacri at 0.6. much closer. >>15504 yeah starting at 0.2 rather than ending at 0.2 probably would have fixed it, but i was hoping to use the shuffle gibberish to provoke more interesting composition like >>15416 worked okay, just broke on the highres fix pass. time for more experimenting i guess.
>>15507 >Use only-mid-control on high-res. fix Ah yeah, I was thinking this might help too but didn't want to suggest it blindly. That setting was actually just added back again today but under a different name, so now it shows twice in the UI for those of us with an old install. I should make a PR really quick to fix that. >but i was hoping to use the shuffle gibberish to provoke more interesting composition Agreeable point too, it is fun to see what it'll give you at full strength and all steps.
>>15508 nvm, those are actually two separate settings, got really confused there for a sec. The new one applies only mid-control during the whole inference process.
Thank fuck for this, I didn't want to inpaint the foot but I didn't want to change the seed/composition either. Although she's upset that I took 10 mins to find a good seed instead of just inpainting.
>>15507 >much closer in my experience it should be almost identical with correct seed and settings (againt, xformers crap aside). I use img2img upscale most of the time because hires fixing all images takes too long on my potato
>takorin lora can't do reisen WHY
>>15510 Sorry if this is a dumb question but did you post your Shoebill LoRA anywhere? Searched for it and all I found was a forum post (maybe that's you)
>>15514 >forum post I'm not on any SD-related forums, wtf Also no, I haven't posted it anywhere mainly because it's somehow both overfit and underfit garbage that I've been wanting to retrain since day 1, wanted to build a 3090 rig for that but I haven't had the energy to do anything these past few weeks
>>15515 >I haven't had the energy to do anything these past few weeks I'll try to get it done before SynthV Teto's release on the 27th
>>15515 >>15516 Cheers, take your time anon
(701.67 KB 768x768 catbox_qayq40.png)

>>15516 Damn I've needed to redo that Teto LoRA for months but I have no computer at all, almost wish I'd never left
>>15518 I mean, yeah. But you know, Touhou artist and whatnot. Kinda expected it to work ootb. It's probably the AOM2-corneo model the takorin baker recommended, I've been using that one since it seems to work better than b64v3 for this but it HATES certain LoRAs and 2hus. Why do you still use underscores btw?
>>15517 Been taking WAY too much time tbh, took a 2 week break when I first got all the parts, then I had to RMA the RAM due to DDR5 shenanigans and now it's been 3 weeks since I had everything ready again. Maybe I should chug some monster, a bottle of 5 hour energy and a dozen vitamin D pills all at once and see if that'll fix my lack of energy >>15519 Maybe wait a few months, need art of her new design. I love the design itself but the mouth and especially the eyes are way too creepy
>>15522 prolly xformers
>>15523 That's WAY too much to be xformers' fault, I'm thinking different series cards
>>15523 >>15524 If it's not hardware differences it could be easynegative, I grabbed this one off of civitai since I've never needed it until now. Could you hand me yours?
also holy shit I just noticed how overfit the takorin lora is now that I'm not doing Shoebill through a LoRA or Momiji (yeah yeah I'm one to talk) Even when I tone it down to 0.3 it still gives Parsee's clothes the fucking weimerica flag's colors, it ain't june already
>>15432 The growth must be very huge since some collab has the option to use it from phones without killing the tab.
>>15527 you really underestimate the third-world swarms circling anything AI and free
(1.65 MB 1024x1536 catbox_6za30v.png)

(1.99 MB 1024x1536 catbox_g3vrhn.png)

(2.03 MB 1024x1536 catbox_kbt3qy.png)

(2.19 MB 1024x1536 catbox_pf68o9.png)

>>15488 Looks interesting, but I know so little about comfyui that I have no idea if this Is a good tool or not? To be fair, the node system is seemingly really powerful, so if you plan on excluding it altogether I suggest not, but beyond that... I have no clue, I feel like a but more info would be needed. To be entirely fair though, I'd much rather just move away from web based uis in general, I don't like how often they seem to just break on me.
>>15510 how exactly does variation seed work? I always keep it at -1 since I don't have any other seeds to work off of
>>15530 Performs a slerp between the main seed and the variation seed. https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#variations
>>15530 You don't have to touch the seed, just use the slider. When you find a variation you like just recycle the variation seed then highresfix/upscale
>>15528 Yeah, and you underestimate how many thirdie having only phone(even multiple) devices and no laptop/PC,
reupping the tekuho locon that got shoah'd. we really need a site that's less cucked than civitai. https://files.catbox.moe/bv0268.safetensors
>>15529 Yes but it seems that webui's ubiquity is perpetuated by three factors - Able to use remotely - Support for both mobile and desktop - Ease of extension With these three intersecting demands the most realistic way for a lone dev to fulfill them at once is using a web-based framework from my understanding, I don't know of an easy way for example a poser extension to implement the necessary UI natively on both desktop and mobile without two codebases and two development toolchains. Unless you'd want to give up mobile support entirely, which I don't want to do. I started this project partly because it was so annoying to use webui from a phone and wanted to improve the UX, while keeping the power that webui provides
I reread old threads and I found this image. https://8chan.moe/hdg/res/1310.html#1317 I really love this style. Dose anyone know the model and(or) style LoRA? (and if the creator read this, I want to know the prompt of this image.)
>>15536 Hey that's mine. Well I found it at least. So thanks. The style is "maco22" combined with "dowman sayman", actually this image used an old version of the former model before I redid it, turns out it was extremely overfit. I can maybe use the image browser extension to skim for the original but I'm far away from my computer right now, in the meantime here's my mega https://mega.nz/folder/oUokGaJR#OWRSRLgTrr9MrbPzkAFfYA
>>15537 Thank you!! I'll download those. And waiting for the prompt if it's not too much trouble for you XD
(1.23 MB 1024x1024 1682126034673291.png)

(1.10 MB 1024x1024 1682126611825039.png)

>>15538 Was a pain in the rear but I got the original, here you are https://files.catbox.moe/m4ojfe.png
>>15540 Thank you for taking time!
>>15535 Fair enough, that being said have you seen the ux fork of webui? https://github.com/anapnoe/stable-diffusion-webui-ux
>>15542 Isn't this the guy who endlessly shilled his large pull request that had a donation link to him, and refused to remove it?
>>15542 I've seen it but haven't tried it yet, I always assumed it was an unstable fork that he wanted to merge into mainline webui instead of a fork whose development diverged completely I can't set it up right now but I have a feeling that gradio/webui's architecture isn't well suited for bad network connections with how much frontend to backend communication is necessary, ComfyUI was designed so the UI can lose connection and reconnect at any time without losing any settings. And I don't know how compatible it is with mainline webui's extension API being a fork. That being said I'll probably end up switching to it when I get back since it's at least being updated frequently, at this stage voldy's repo is practically abandoned with the complete lack of updates for so long
>>15544 And another thing is ComfyUI has a built-in prompt queue which gradio isn't likely to support for a very long time, that alone would be enough reason to switch development to a different base repo since having a useable queue would have saved me hours of time just waiting for the current prompt to finish generating so I could test out a different set of parameters (I tried that JS hack extension for queueing on top of webui but didn't find it friendly to use)
>>15543 No clue, just thought it looked nice >>15545 Uh sure, I still haven't tried comfyui and it's probably going to be a while before I do, as I don't really want to learn the nodes. Have never touched a node system like that, and it seems annoying to do most of the simple things. It does look like it can do a lot of interesting things with enough tinkering though.
>>15543 That guy also stole most of the code within that pull request
>>15459 is there an alternative to automatic1111's "prompts from text box or file" script? i can't get a1111 to use the prompts from my text file no matter what syntax/examples I've copied from threads or the github page. also would be nice if there's something that allows specifying the denoising strength for each prompt job as well.
>>15546 Yes, I'm in agreement that it's annoying. That's what the UI I'm developing is supposed to solve, a way to never have to bother with the graph after you've configured it to your liking. It'd be like webui but if you could open up the hood and rewire things if you really wanted to, without needing to code anything It's another benefit of ComfyUI, the frontend and backend are well separated so it makes developing custom frontends easier
>>15546 Yeah node UI is a good idea, it's just a shame that the dev is such an unbearable faggot
anything noteworthy happened with the webui or training?
(1.92 MB 1024x1536 catbox_vcbns4.png)

(1.96 MB 1024x1536 catbox_a56ght.png)

>>15552 Regional prompter I guess but I'm too stupid to figure it out, too much cunny juice has fried my brain. https://github.com/hako-mikan/sd-webui-regional-prompter
>>15552 This one released a bit ago. Paper claims it's supposed to be an overall improvement and in my testing so far it does seem to be that way. https://github.com/ashen-sensored/sd_webui_SAG
(2.06 MB 1024x1536 catbox_2zv2q0.png)

(2.03 MB 1024x1536 catbox_z0o842.png)

(2.01 MB 1024x1536 catbox_jv99lb.png)

(1.93 MB 1024x1536 catbox_r0sw86.png)

>>15554 odd, in my experience it only really helps in improving the color, that's about it, for example. first image is without SAG, second is 0.75 guidance, third is 1 guidance, and fourth is 2 guidance, as you can see, it does improve colors a bit, but it's not a meaningful upgrade in any way past that... that being said, it's not detrimental, so you might as well install it, set it, and forget it.
>>15555 Pillars in the background look a lot better to me even at 0.75. 1 and higher might be pushing it too much
>>15557 marginally better, still an improvement, but not a ton. But I didn't notice a slowdown at all so... fuck it, install it and just always use it. literally a set and forget type of thing
>>15552 big controlnet update
>>15554 >made my gens slower no thanks
>>15560 >slower gens on top of the already other slowing extensions yeah, no way fag.
(2.51 MB 1024x1536 catbox_bbywos.png)

(2.54 MB 1024x1536 catbox_583fog.png)

(2.20 MB 1024x1536 catbox_k33qyg.png)

(2.18 MB 1024x1536 catbox_crm02n.png)

>>15554 honestly pretty nice. roughly 40% speed hit on a 2080 ti, 19 seconds vs 33 seconds each.
(2.30 MB 1024x1536 catbox_5wwi8d.png)

(2.30 MB 1024x1536 catbox_6q8mkr.png)

(2.13 MB 1024x1536 catbox_tkaw1m.png)

(2.04 MB 1024x1536 catbox_1zkn20.png)

>>15562 just for fun >vanilla >+ SAG >+ dynamic thresholding >+ controlnet shuffle 1.7 weight 0.0-0.2 steps with some watercolor swirls lora mix goes from non-functional to merely crappy and overwrought
>>15546 >I don't really want to learn the nodes You can load a generation from the gradio webui and it'll recreate a workflow based on that. My only issue with it is that it doesn't seem to import the seed when you do that/doesn't set the seed to static.
>>15564 I mean sure, that's a cool feature, but the problem of me still not knowing how it works is still there. the workflow might be in front of me, but It's likely that it won't be enough on it's own to explain how to use the program. If i'm going to use comfy, I'm going to learn how it works enough to do the fun stuff I've seen people make, and that's a bit too much of a time sink for me right now
>>15565 >I mean sure, that's a cool feature, but the problem of me still not knowing how it works is still there. Which is why you should load a gen and look at the nodes. >but It's likely that it won't be enough on it's own to explain how to use the program Come on dude, it's not rocket science and it's not Houdini nodes levels of autism either. Takes literally five minutes to figure out what the nodes do and how you should connect them based on previous gens. I'm the "failed high school maths" kind of brainlet and I still figured it out.
>>15566 uh, sure. maybe another time, currently fighting with PySide2 to make a UI
>>15567 >currently fighting with PySide2 to make a UI >can code >in the clusterfuck that is python no less >still thinks nodes are too hard to understand without rtfm
>>15568 less that I think they are too hard to understand, but too tedious to understand
>>15569 >tedious >coding I'm just gonna give up.
>>15570 I'd say there's a pretty big difference in coding vs genning. a tedious way to gen vs a clear cut easy way to gen is a pretty hard sell. But regardless, I was already going to check it out, eventually. Just gotta see if it's possible to get the wildcards system I normally use working in comfy.
>>15571 here you know what, I'll get it installed now so at the very least I can't use not having it installed as an excuse to not try it
>>15572 >I can't use not having it installed as an excuse to not try it I'm not here to shill it but people making all sorts of paper-thin excuses to not do something just pisses me off to no end.
>>15573 hey, in this case I'm removing the option.Genuinely speaking though, I just hadn't had the time to look into it up to this point, and honestly hadn't felt the need to. so it kept being on the back burner.
Guess the artist challenge (impossible)
(3.26 MB 1280x1920 catbox_ci6w81.png)

>>15576 Must be sakimichan.
If I removed all the text, cropped and tagged doujinshi art would anyone be interested in cooking a Lora for a doujinshi style. Every time I find a style lora for a popular doujin artist it's cooked with only the 40 or so colored images on booru and looks nothing like their actual work.
(3.04 MB 1920x1280 catbox_fu2ddn.png)

>>15579 Not even gonna bother removing the logo because it's tagged and doesn't show up most of the time anyway
I'm throwing shit at the wall to see what sticks. Some of it sticks pretty nicely and I'd like to know why but on second thought I'm not gonna investigate least it starts to slide down and fall off. And by that I meant I'm not sure if this mix is even working as intended or if some of these even do anything but I like what I got.
>>15582 When I saw the Fiz LoRA I knew I had to try mixing it, and as suspected, it gave an interesting effect to faces, but it's a fine balance sometimes so it doesn't look uncanny. I usually go for like a 0.2 focus.
>>15584 Yeah, it is. Even at low weights it's pretty damn strong. But it's damn good too. Anyway, this might be one of my favorite gens ever
>>15582 try LORA block merging for more control over the details
>>15586 >LORA block merging Is there a guide? I'm probably too dumb for that
>>15583 Shiiit, they probably changed how that consent dialog response gets saved to the browser cookies again. That's what broke it last time. I'll try to take a look at it soon but no guarantees.
>>15587 I don't think there is actually, I have used it to just strip baked in character styles from a LORA so I can give them another one using a different style LORA using "OUTD" I mean here is the link to the extension if youre interested. https://github.com/hako-mikan/sd-webui-lora-block-weight
>>15591 gonna add I think the example image does a decent job of letting you know what block corresponds to what a LORA does
(1.38 KB 278x36 msedge_BDpmZmTmil.png)

>>15592 Sounds useful since my soon to be rebaked Shoebill LoRA will most likely still have a heavy bias towards guchico's style. All I need to do is use this as the weight, right?
>>15593 yeah "OUTD" are the ones that correspond to character traits usually, sometimes "MIDD" as well number on the left affect style/colors & pose from my findings.
(3.40 MB 1280x1920 catbox_l5udf7.png)

(3.23 MB 1280x1920 catbox_dyovr8.png)

(2.74 MB 1280x1920 catbox_tuazya.png)

(2.92 MB 1280x1920 catbox_lxjuy1.png)

>>15584 Nice I never really thought to mix fiz rot like that
(2.81 MB 1920x1280 catbox_c9ytqi.png)

(2.99 MB 1920x1280 catbox_h8p3dz.png)

(2.88 MB 1280x1920 catbox_j6ow79.png)

(2.78 MB 1280x1920 catbox_kk6703.png)

finally stopped being lazy and picked up controlnet. this QoL upgrade is insanely fun
So that SAG extension is pretty cool, I like how it makes stuff look sharper
>>15603 Sadly this one really hates my LoRA, tried to tone it down to 0.8 but it's still very wonky
(2.71 MB 1280x1920 catbox_9lqhz3.png)

(2.39 MB 1280x1920 catbox_kuul7t.png)

(2.64 MB 1280x1920 catbox_oq8yzi.png)

>>15604 It seems to be having the same effect as the S16xue lora which is about what I expected. More realistic styles tend to overfit pretty easily.
>>15583 Pushed an update, could you try it again?
>>15606 works great now. thanks chief
(2.24 MB 1024x1536 catbox_ge57w9.png)

(2.16 MB 1024x1536 catbox_u737t4.png)

(2.15 MB 1024x1536 catbox_2nwb8t.png)

>vanilla >dynamic thresholding >self attention guidance slowly discovering the final solution to the breakfast sausage hands problem
>>15608 >alkemanubis there's your shitty hands culprit
(1.79 MB 1024x1536 00223-3250813174.png)

(1.93 MB 1024x1536 00010-152811610.png)

(1.78 MB 1024x1536 00011-705659729.png)

(1.59 MB 1024x1536 00185-3879797652.png)

https://mega.nz/folder/ctl1FYoK#BYlpywutnH4psbEcxsbLLg added sakura yuki (clochette) style
(1.72 MB 1024x1536 00017-3919829186.png)

(1.53 MB 1024x1536 00013-1369748150.png)

(1.78 MB 1024x1536 00076-2629809589.png)

(2.43 MB 1024x1536 00042-3706432232.png)

>>15609 >0.0 just forgot to remove it after testing
>>15611 Oh nice it's the shortstack artist
So it seems like >reddit is shilling the Vladmantic webui, more reason to skip it https://www.reddit.com/r/StableDiffusion/comments/12wgow1/vladmantic_and_anapnoe_collaborating_on_the_best/
(3.51 MB 1280x1920 catbox_7yolqu.png)

(3.52 MB 1280x1920 catbox_sgudgw.png)

A healthy diet
>>15614 and because jewtube shills comfy and twatter shills webui guess that means we'll just have to give up on using stable diffusion forever tbh from a realistic standpoint those two coders cooperating on a fork of webui that is actually maintained is all that matters. just want updates and someone who's not a complete retard like voldy so they wont dump a steaming pile of shitty code on master then piss off for a month to binge on dota
>>15616 lmao the person vlad is working with is the same one that made a shitty css stylesheet for the webui and tried sneaking in a donate here button to be committed. auto turned him down for good reason
>>15617 > the same one that made a shitty css stylesheet for the webui and tried sneaking in a donate here button ok thats more than enough reason to not use this one
>>15618 why do github coders even have lame ass autistic drama like this, it's as bad as vtuber drama
>>15610 cute
>>15617 i personally don't care unless he sells his ass to a corporation like what happened with audacity, then he's an unforgivable cunt im just tired of having to maintain a custom fork with two dozen patches that's gonna get outdated in a month and complete lack of traction on fixing gradio's garbage (shitty dropdowns, freezing, unresponsive ui)
>>15617 >retard that can't code for shit and is the source of a lot of AUTO's UI bugs >smoothbrain who stole code and slapped a donate button on it What a duo!
>locons work even at epoch 1 what in tarnation
Looked at a /g/ thread talking about what gpus people have and one anon got one of the Yeston chink waifu GPUs, did any anon here even get their hands on one of their products? I would be more inclined to buy one of their products if they took their chink anime artists skills and made silver haired lolis their mascots
>>15627 newegg had a really good deal on yeston 3080s a while back that i missed. apparently pretty crappy cards.
Can you use SAG for img2img?
>>15628 rip, well if they even have a genshin collab and make a Nahida gpu I'm giving my soul up and buying it kek
>>15447 >oyakodon https://files.catbox.moe/dkaa5n.png Bit late to this, but you can use this as a template. No idea if this is the best way to do it, but it works for me (most of the time).
>>15631 Forgot to mention that the numbers after ADDCOL are not necessary and only add noise, I just have them there for easy reference to which block I'm putting prompts in.
(1.75 MB 1024x1536 catbox_94gt5h.png)

(1.88 MB 1024x1536 catbox_cxd9uw.png)

>>15631 Thank you kind sir, I will study this.
(1.84 MB 1024x1536 catbox_1o56ep.png)

(1.86 MB 1024x1536 catbox_b6rss7.png)

(1.86 MB 1024x1536 catbox_db437n.png)

(1.92 MB 1024x1536 catbox_yrng2u.png)

Trying out padoruu lora
(1.94 MB 1024x1536 catbox_j7lvdh.png)

(1.85 MB 1024x1536 catbox_s41fqj.png)

(2.00 MB 1024x1536 catbox_u7l5vv.png)

(2.06 MB 1024x1536 catbox_t7ifzk.png)

>>15635 first image uploaded correctly though
>>15636 Yea idk. That was attempt 4 out of 5. The rest completely failed.
>>15627 are you talking about this? My friend got one and he let me keep the box. The GPU is apparently scented
>>15639 >scented the chinks went far with their waifu gpus huh
>>15616 >and because jewtube shills comfy and twatter shills webui guess that means we'll just have to give up on using stable diffusion forever >so they wont dump a steaming pile of shitty code on master then piss off for a month to binge on dota absolutely retarded faggot logic, kill yourself immediately as bad as voldy's ui is, vlad's fork is genuinely cancerous and you'd know just by looking at it and its baked-in extensions for longer than half a second
(2.57 MB 1280x1920 catbox_kpytb8.png)

(2.54 MB 1280x1920 catbox_6hsxb8.png)

(2.31 MB 1280x1920 catbox_hhpjfz.png)

Ataruman. Surprised I haven't seen this one yet https://mega.nz/folder/2FlygZLC#ZsBLBv4Py3zLWHOejvF2EA/folder/mIkl3IpD
(1.26 MB 1152x1152 catbox_jpv7iv.png)

(1.49 MB 960x1344 catbox_srbewr.png)

(1.48 MB 1152x1152 catbox_slqlju.png)

(1.56 MB 960x1152 catbox_bqnhh2.png)

added kionant, can probably prompt gardevoir but I didn't try lol https://mega.nz/folder/OoYWzR6L#psN69wnC2ljJ9OQS2FDHoQ/folder/PghTGDzA
>>15642 i'm surprised too. thanks~
>>15643 >can probably prompt gardevoir but I didn't try this looks like a job for me
(1.63 MB 1024x1536 catbox_01vglr.png)

(1.68 MB 1024x1536 catbox_k7b7ld.png)

(1.79 MB 1024x1536 catbox_jwara0.png)

>>15642 >>15643 Very nice, thanks(this prompt is gonna kill me btw).
>>15647 see? i knew that telling you to make them pregnant was an excellent idea
currently working on something which made me wonder is there a rance lora?
>>15649 just prompt kazuma with an evil grin
>>15648 Yes it was an excellent idea but not so excellent for my health.
>>15650 no I mean a style lora in the style of the modern rance games
Heard AM5 processors were starting to blow up, but apparently it's due to some overvolting that some overclocking profiles "EXPO" that is causing it to make the voltage go above safe levels for the CPUs causing them to "blow up" the motherboard and processor, if any of you are running the x3d or normal chips of this AMD gen make sure you SOC voltage is set below by .1 of the auto preset in your bios until all the manufactures release a bios update that solves these overvoltage issues
(2.08 MB 1024x1536 catbox_e3azdb.png)

(2.07 MB 1024x1536 catbox_k2ojb1.png)

(2.09 MB 1024x1536 catbox_bj0mso.png)

Here is a set of images I was working on. I think in the end they turned out pretty good. I didn't to any fixing, the time I spent was on getting the right kind of gen in the first place while keeping as much of the details as I could from the other variations. Took a bit because I'm not used to it.
>>15653 no no no no NO NO NO NO NO NO NO NO
>>15653 >asus pushes a bios update and also hides all their old bios uh oh stinky
>>15653 >>15655 >all on asus x670s NO NO NO NO NO NO NO NO NO >only X3Ds phew i'm safe
>>15652 There were Rance hypernetworks very early on made when NAI first leaked, (if I can find them in my drive I'll probably upload them) but I dont thnk theres Rance style LORAS
>>15655 >>15656 kek I didn't notice the update would be that fast but this also happened with someone else but on a gigabyte motherboard with a 7900, so yeah it seems to be an issue with other MB manufactures that might also silently release their bios updates >>15657 https://youtu.be/arDqhxM8Wog
>>15658 welp working on one right now just wanted something to compare
(924.08 KB 498x277 AQUASCREM.gif)

>>15659 YOU MOTHERFUCKER I WANTED TO BUILD MY 7950X X670E-E RIG TOMORROW AFTER NEARLY 3 MONTHS AND NOW YOU COME UP WITH THIS FUCKING FUCK WHY'D YOU HAVE TO DO THIS
>>15660 Yeah they are very early ones so anything you bake will be 5x better than those.
>>15661 it should be fine, there's already enough of drama posting on youtube and >reddit so the manufactures will just fix the overclocking profiles for their motherboards not too long from now, just run the EXPO profile and manually adjust the SOC Voltage and you're good to go
>>15654 Cute
>>15663 From what I'm reading it seems like ASUS released a fix already, do I just boot, update the bios, reboot, apply EXPO and then lower the voltage by 0.1? Wasn't really planning to upgrade the BIOS but oh well
>>15665 gl don't brick your machine t. also has an aSUS but has intel and doesn't overclock.
>>15665 if the bios update to fix the voltage issue has already been released then it should be fine, I mean you can still lower the SOC voltage if you just want to be safe but it's mostly the X3D chips that are mostly likely to get destroyed by the voltage issue so there shouldn't be such a huge problem on your side.
>>15662 got it will play around a bit and then upload so that other anons can call it shit or praise it
(1.88 MB 1152x1152 catbox_844v62.png)

(2.12 MB 960x1344 catbox_el8rnm.png)

(1.58 MB 1152x1152 catbox_t2i7h4.png)

(1.46 MB 960x1152 catbox_i0sjm2.png)

(3.65 MB 1536x2304 catbox_7vvvym.png)

>>15671 fuckin whew. catbox?
(1.32 MB 1024x1536 catbox_oms29d.png)

>>15673 >>15674 nice, thank you
>>15676 Trying it out, it's quite a bit better but still not there. What's the syntax for the lora block weight thing? <loraname:weight:preset>?
(3.00 MB 1280x1920 catbox_nvh9zv.png)

(3.08 MB 1280x1920 catbox_kpqtw2.png)

(3.05 MB 1280x1920 catbox_cbbvwg.png)

I just realized I messed up my folder names in my training image folder, gonna have to retrain both
>>15677 yes, replace "preset" with OUTS, OUTD, INALL, etc. or your own numbers
>>15679 OUTD doesn't really do it, I guess Guchico's style is just part of my LoRA at the moment. But hey, at least it's ever so slightly more compatible with things like Fishine.
(2.55 MB 1920x1118 catbox_exh6d6.png)

>>15676 Alright should be last update fixed the folders and retrained
>>15681 You only updated the DIM16 one right?
Lia hates jannies!
>>15683 The correct answer was Dragon Quest 8 btw.
Just starting out with loras. Are there any settings in the script at the bottom of this "other lora training guide" that need changing like epochs or something? https://rentry.org/59xed3
>>15684 I figured. I've only played a bit of DQ 1, but plan to play all/most of them in order.
(2.94 MB 1280x1920 catbox_hvpiae.png)

>>15682 Both. I deleted the old ones but I kept the names the same. Not sure which one is better at this point so I just put 16 in the extras folder
>>15686 11 plays and looks real nice but the story is absolute dogshit so I just dropped it. The only other game I dropped due to the story was Lightning Returns. Here's some stuff that would grant me a few weeks of jailtime on 4kikes. >>15585 >>15271 >>15277
(2.04 MB 1024x1536 catbox_fnh4m2.png)

(1.91 MB 1024x1536 catbox_22bnwv.png)

(2.05 MB 1024x1536 catbox_ywdtj3.png)

(2.01 MB 1024x1536 catbox_f2u8sh.png)

here, have a fox girl
(1.97 MB 1024x1536 catbox_wmbmr2.png)

(2.16 MB 1024x1536 catbox_xctno9.png)

(1.82 MB 1024x1536 catbox_iu9w1u.png)

thank you once again to whoever trained the soichiru yamamoto lora as for my experiments earlier with using splotchy noise as a picture for a shuffle controlnet to get more interesting compositions, i've found that depth works much better
(2.79 MB 1280x1920 catbox_ytopx3.png)

Someone made a fork of composable lora with some interesting features but I still can't get it to work https://github.com/a2569875/stable-diffusion-webui-composable-lora
>>15687 Style works now but both versions fuck with my Shoebill's consistency (DIM8 one works better which is kinda weird) Guess it's time for me to just fucking retrain this thing. Other than that it works real nice, thanks.
what's an epoch, what's a repeat, what's a bucket, and why do these exist? guides give settings, but don't explain wtf these are
>>15693 >epoch One full iteration over your training set (every image seen at least once, though some might be seen more due to repeats) >repeat How many times per epoch a given image is shown. Increasing the repeats on images basically makes the model train on them more than other images. A common use case is adding some repeats for very high quality images or images that display some sort of scenario that you want to train on but is rare otherwise (e.g. an unusual pose, costume, or angle). The other difference between repeats and just more epochs is that most learning rate schedules are based on epochs. >bucket To take advantage of the parallel processing of GPUs, computations must be vectorized. E.g., turned into a single matrix operation or several matrix operations of the same size. The long and short of this is that it is very inefficient to have every single image in a training set be a unique size. Therefore, images are resized into certain "buckets" of a given size e.g. 512x512, 448x576 etc. Images are resized to the nearest bucket in terms of their aspect ratio (length vs width ratio), with any excess being trimmed (usually randomly). You can control the base size you are training on (do we start at 512x512 or 768x768? Bigger allows more detail but takes more vram) as well as the maximum width/length of a bucket (how extreme the aspect ratio can be).
(2.77 MB 1280x1920 catbox_d8ulw6.png)

(2.66 MB 1280x1920 catbox_nluc35.png)

>>15695 man that's my default hires fix resolution
>>15698 same, and I'm on a fucking 2060 with tiled vae I can even do more than batch 1 at that size
>>15698 >>15699 i gen at 512x768 and 2.5X is the best i can do but the .5 fucks with the gens a bit so i dial it back to 2x
>>15695 "add jeremy clarkson to the prompt" "wtf no" "do it" >it improved the eyes and the head wings are more on-model now what the fuck
>>15697 UOOOH PREGNANT LOLI BELLY, I WANT TO MAKE A LOLI PREGNANT EROTIC UOOOH ToT ToT ToT
>>15702 Yes.
(82.76 KB 512x512 GigaChad.jpg)

>>15703 >Yes.
>>15705 oh! oyakodon so soon
>>15706 One of them ends up getting bleached very often, sad times. Making some good progress though.
>>15707 shit happens next gen gonna be a chocolate parfait with a lot of cream
>>15709 >booba of the lake, what is your wisdom
>>15702 gonna need a meme lora for this
reminder to not feed the feral lolis, doing so can make them dependent on you and they will have difficulties surviving in the wild
>>15712 Cute, just give them a loving home. Also I'm getting stronger with these new powers...
Jesus that new update for SD-Scripts that Kohya rolled out >Naming of LoRA The LoRA supported by train_network.py has been named to avoid confusion. The documentation has been updated. The following are the names of LoRA types in this repository. LoRA-LierLa : (LoRA for Li n e a r La yers) LoRA for Linear layers and Conv2d layers with 1x1 kernel LoRA-C3Lier : (LoRA for C olutional layers with 3 x3 Kernel and Li n e a r layers) In addition to 1., LoRA for Conv2d layers with 3x3 kernel LoRA-LierLa is the default LoRA type for train_network.py (without conv_dim network arg). LoRA-LierLa can be used with our extension for AUTOMATIC1111's Web UI, or with the built-in LoRA feature of the Web UI. To use LoRA-C3Liar with Web UI, please use our extension.
I can't believe lora makers are fucking dead
(1.73 MB 1024x1536 00053-54257722.png)

(1.62 MB 1024x1536 00016-2224148577.png)

(1.79 MB 1024x1536 00009-1524530704.png)

(1.55 MB 1024x1536 00017-2714547791.png)

(1.57 MB 1024x1536 00017-3362734109.png)

(1.75 MB 1024x1536 00234-3660509702.png)

(1.91 MB 1024x1536 00194-3604778292.png)

(2.57 MB 1280x1920 00092-837754942.png)

so apparently people are saying vladmandic is making better outputs because of this, snake oil? >Vlad provides more optimization options than auto. Differences in image detail are likely due to these options. Vlad is updating very quickly. I submitted 2 issues related to extension compatibility and both were fixed in less than a day.
(2.71 MB 1280x1920 catbox_dpv0ty.png)

imminent penetration
>>15720 It doesn't, it just exposes the settings in the UI instead of CLI flags. The defaults differ between voldy's so by default it will also produce slightly different results.
>>15720 >>15722 Also, that screenshot doesn't help, because you can't tell what optimizations he's using for voldy. What he's using for Vladmandic, scaled-dot-product with memory attention, is equivalent to using xformers, so it's already a flawed comparison.
Anyone still taking requests? Honkai Star Rail comes out today and I'd love to request a clara lora if possible
>>15724 any lolis that will make me pick up this game?
>>15726 >train in intro I know Bronya is in this game but did Kiana's actions at the end of the "first part" of Honkai Impact 3 give birth to this game or something
>>15726 Clara is cute but she's currently a pretty average unit that can swing either to mediocre or godlike depending on what they do with the game Bailu is one of the best units right now but I don't see her lasting long in the meta unless they simply don't release many healers My hopes aren't high because it feels like the Honkai team really fucking hate loli given their treatment of Theresa but I'm hoping it's good
Huh, who's that new maid? I'm not sure how I'm supposed to stop the AI from making one of them white.
4chan /hdg/ is saying repeats are bad now for LORA training, thoughts?
>>15724 well seems like someone already made a LORA https://www.pixiv.net/en/artworks/107512541 I'll do it because she doesn't have a lot of images right now so it'll be less of a hassle to deal with tagging and filtering
>>15724 sorry bro, I'll be rerolling star rail myself so I won't have time to train cunny loras
>>15729 your own fault maid anon too much breeding of dark skinned beauties makes them whiter each generation if you bleach them should have seen this coming
>>15733 It's over...
>>15715 can i get a catbox for the first one? that's a really cute style.
>>15694 >Increasing the repeats on images basically makes the model train on them more than other images. define "other images". i thought repeats were applied equally to all images in your training set
>>15699 are there any drawbacks you've found using the tiled vae?
(2.75 MB 1024x1536 catbox_8qqvpz.png)

>>15736 thanks dude
I'm just going retard mode on the MBW values for LORAs, anyone that actually knows what they're doing has tips?
(1.22 MB 960x1344 catbox_dun7ii.png)

(1.29 MB 960x1344 catbox_bfuwvi.png)

(1.29 MB 960x1344 catbox_4qggmo.png)

>>15739 ur welcome have this even softer mix I just concocted
(2.75 MB 1280x1920 catbox_kts7wk.png)

RuntimeError: The size of tensor a (40) must match the size of tensor b (80) at non-singleton dimension 3 Well I was hoping to use the metata locon but I guess I'll to stick to the lora for now. And stop procrastinating on making a Roll lora
tried based v3 >>15553 seems to do the much of the same latent couple does? seems like a lot of detail is lost when it's used but maybe that's just me doing stuff incorrectly. >>15554 i think i'll be using this. i like the contrast and seems it's making stuff more coherent? or not?
>>15744 it does a great job of cleaning up edges sometimes, others it just adds contrast and waiting
>>15744 I never got latent couple to give different girls the correct hairstyle, breast size etc but could have been a skill issue on my side. Now I sleep.
The whole talk about more epochs than repeats on concept folders making better LORAs might make me rebake everything, pain
>>15737 They'd be applied to all of the same in the same folder using the Repeat#_Concept format that most scripts use. There's basically no point (or harm) in using repeats if you're repeating everything the same. You're basically just making each epoch into repeat amount number of epochs. I guess it has some side effects on shuffling and LR schedule - very minor.
Anyone have a link to a good defloration lora, or know tag combos for it? Can't seem to get it to work. There's no way you degenerates don't have this knowledge.
>>15743 >They're all yours anon] welp time to start looking for another job I guess
Star Rail rerolling is pain, I miss when MIhoyo's games were simpler to reroll wtih now I have to make multiple emails
>>15701 Where did you hear that from? And yeah it does pretty much work, interesting.
(2.28 MB 1024x1536 catbox_amcdwp.png)

I made this as part of some challenge I was doing. ended up taking a lot of effort because, at it turns out, having two BA character lora's running at the same time really causes the halo to clash. I think it turned out pretty good, but it required a lot of halo inpainting... so how do you think I did? oh... also, prompt is: shun \(small\) \(blue archive\), 1girl, black hair, long hair, bangs, twintails, hair ribbon, (tiger ears:1.1), tiger girl, halo, green eyes, flat chest, (china dress:1.2), (vertical-striped dress:1.2), (frilled skirt:1.05), off-shoulder, jacket, (cameltoe:1.2), standing, pussy juice <lyco:shun:1>, <lyco:kokonaBlueArchive_v1:0.7> Negative prompt: EasyNegative, censored, sash, Steps: 26, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 426701150, Size: 512x768, Model hash: 98a1428d4c, Model: based64_v3, Denoising strength: 0.55, Clip skip: 2, ENSD: 31337, Hires upscale: 2, Hires upscaler: Latent (nearest) because I inpainted a bunch, so it overwrote the original prompt in the metadata
(269.64 KB 373x910 Discord_VKGyj5XTFB.png)

(532.38 KB 425x984 Discord_cVoz1I7b3w.png)

(479.93 KB 1296x714 Discord_RU47NeJQg1.png)

>>15753 It's more or less a conversation I had with a friend it actually does work
>>15755 depending on the prompt I've had success with adding "charlotte gainsbourg". wonder if other names will have positive effects.
>>15697 catbox?
>>15759 oh no I brainfarted again... coffee save me...
>>15756 i will now add jeremy clarkson to all my prompts for extra SPEED and POWER
sup guys, I've been absent since the day of the last voldy update iirc. what did I miss? did XTI ever become a thing? Is this dylora stuff worth looking into?
Oh, my favorite reference, why thank you SD.
>>15763 pillar person... pillar men... IS THAT A FUCKING JOJO REFERENCE?
Sharing bratty maids with anons, the choco ones are mine though.
>>15759 >catbox_tsoxyz I'm so confused how this prompt works for you when the Regional Prompter parameters are so messed up.
>>15766 It just werks, well I literally just used that anons catboxes and changed the prompt but I understand how the shit works now so future gens won't have messed up parameters.
(1.99 MB 1024x1536 catbox_pb8qj8.png)

(536.45 KB 512x768 catbox_fzrzda.png)

(1.76 MB 1024x1536 catbox_v0ih75.png)

(1.96 MB 1024x1536 catbox_nscv2f.png)

>>15767 Cool. I'm trying it out more now since I really only did when it was first released, and it seems to give me mixed results, but when it Just Works like the last pic, it's great. Looks like it's soon about to get an upgrade too (this was only merged on the dev branch): https://github.com/hako-mikan/sd-webui-regional-prompter/pull/66
>>15759 holy fuck that loli in the last pic is about to gut punch the preggo
Fucking less concept repeats and more epochs is bullshit, my LORAs that balance the repeats and epochs look better than just 1 repeat and 30+ epochs, fuck 4chan for saying shit that doesn't work
>>15771 1 repeat is shitposting
>>15772 It really depends, more epochs can work if you have a small dataset because trying to balance repeats with less than 30 images leads to over-frying more easily maybe that's why those type of posts are floating around because I expect most 4chan anons to not collect more than 50 images
>>15773 Absolutely this. If you have a ton of repeats on a small dataset, you are likely going to have some pretty serious burn in. I'm much like the 4chan crowd in terms of dataset, I don't collect big datasets, I collect high quality datasets and I've 100% seen better results with lower repeats because I don't have to worry about the batches being the same image back to back.
>>15773 >>15775 Is this because the order is only shuffled between epochs and not repeats? I haven't looked at the code yet but that would make sense. From my understanding I thought repeats were only ever useful if you have a huge, varied datasets, like those used to train LLMs (see the LLaMa paper). When I get a chance to look at the code later I can probably answer this for myself.
>>15776 From what I remember they are shuffled per epoch not per repeat. I think your assumption of larger datasets enjoying repeats is correct.
>>15776 Why would repeats even be used globally? The only real use to repeats (which is already tremendous) is augmenting some underrepresented parts of the dataset
>>15778 Yeah, that's why I mentioned LLMs, specifically LLaMa. That is essentially the only useful factor of repeats from my understanding.
(2.09 MB 1280x1920 catbox_nqx0f7.png)

>>15780 Did someone already make a Bailu LORA?
(1.69 MB 1200x1600 catbox_8bz5nl.png)

(2.51 MB 1920x1280 catbox_361ilg.png)

At long last Nendoroid lora https://mega.nz/folder/2FlygZLC#ZsBLBv4Py3zLWHOejvF2EA/folder/aB11QC4T With a dataset of 10k images only about 1/7 of the images were actually used. Oh well. >>15781 I made one almost exactly a month ago and just gave it a minor update https://mega.nz/folder/2FlygZLC#ZsBLBv4Py3zLWHOejvF2EA/folder/uc0hHQQJ
Not sure if this is good enough for the anon that wanted a Clara LORA, I made a few versions and this one was the one that worked the best with the 40-50 image dataset I had, I'll upload it to Civitai and then post the link here, just need to test out how flexible it is with changing her looks
>>15773 what's a good repeat/epoch number for small datasets versus large datasets assuming that the images are around 1000x500px to 1500x1000px-ish? it's like pulling teeth trying to get an answer on 4chan
>>15784 I just try to focus on how many steps I'm going to end up with depending on the optimizer I use, adamw will be good with 1000+ while lion is good with 1000 below, batch 2 ofc you have to double that if you're doing batch 1 but otherwise I try staying within the range of 8 epochs and if I can use 3-5 epochs with a 50+ dataset with less than 10 repeats for the dataset I am using that will work for me, I'll do the 1 repeat and 30+ epoch method that 4chan is shilling when I have less than 30 images, it works sometimes
apparently a x2 4090 user came back with this type of post on 4chan >Trained 180~ character LoCOns and 5 concepts using dadapt at 640x640. It took ~16-17 hours to complete in 2x4090 IoI (aiming for ~1000 steps with batch size 7 for each GPU, and 2000-3000 steps for concepts) Though got pretty nice results, I tried LoHAs and seems to be inferior, tried just locons and probably used bad settings, since didn't got good results either. Also tried just dadapt and got less detailts than 8bit adam with my settings. It seems dadapt + locon can get more details and you don't have to tweak lr, text lr or unet lr. Finally, for characters, LoRAs are still probably better, but I feel I get superior for concepts with LoCOns vs LoRAs. For styles I have no idea, but maybe LoCOn is better?
>>15786 >For d-adapt used 1 lr, text lr and unet lr. For locon, used "conv_dim=64" "conv_alpha=1", and net dim 64, network alpha 64. That last thing is pretty important, since it seems that conv alpha is different that network alpha, and as we know, dadapt as you go lower on the alpha compared to the dim, the real LR is higher (so more possible fried results)
Why can't I recreate an image made with a locon? It always looks completely different
(2.75 MB 1280x1920 00030-3503921839.png)

(2.30 MB 1280x1920 00029-3503921838.png)

(2.87 MB 1280x1920 00034-1405232245.png)

(2.83 MB 1280x1920 00032-1405232243.png)

witches. also background of the third one kek
>>15788 can you catbox the image and your result plus share the locon? it's probably a discrepancy in your general SD configuration.
(2.83 MB 1600x1600 catbox_koikas.png)

Beware the forest's mushrooms
(1.04 MB 1216x812 msedge_rqDBUFTxIH.png)

time to make a lora out of nearly zero good images
>>15790 https://files.catbox.moe/kjddc1.png https://files.catbox.moe/smfg68.png https://mega.nz/file/e7phmQRQ#LyR0R07P1GLRSvKBrURMeBpo2rmkV9TbvsBJye-daz0 First is what I can't recreate for some reason. I get the second consistently now. Not sure what I'm missing
(1.89 MB 1024x1536 catbox_gmuzk8.png)

I forked ddtailer a bit ago, the PR of it that added the ability to put a positive and negative prompt and cfg scale, fixed up the install reqs, and started using it. this time it outdid itself.
(1.72 MB 1024x1536 catbox_kg3o4y.png)

(1.57 MB 1024x1536 catbox_ur7ft2.png)

(1.64 MB 1024x1536 catbox_mq5e6e.png)

(1.81 MB 1024x1536 catbox_q1b3my.png)

here's some demon girl waifu OC from me. If you want to know, her name is Rhyle
>this is my oh sea and she's named xD oh boy we're already at that point in the board's lifecycle, huh
>>15797 I've had her for a long time, just never posted her here. I happened to be genning her today and thought, fuck it why not, it's not like I'm the first person on this board to have OCs
>>15797 >oh boy we're already at that point in the board's lifecycle, huh literally who cares
not SD-related but give me literally one good reason NOT to buy it
>>15799 three months until this board is just a handful of prison-gay retards flirting with each others' characters like the /b/ threads
>>15801 >three months until this board is just a handful of prison-gay retards flirting with each others' characters like the /b/ threads aren't you a fun one
(291.62 KB 1683x2048 depressed_miku_2.jpg)

>>15800 >AI took my job
>>15803 blame yamaha
At least no one is posting like schzio diffusion general here, I noticed almost all the AI threads on 4chan are like that now. I don't care if someone posts their OC, as long as if they don't post like a retarded tripfag I don't care
>>15805 rather have retarded roleplaying than grannydiffusion tbh
(2.19 MB 1344x1328 00020-713707424.png)

>>15672 Can you post the parameters for that image if you still have them? It looks like the copy on catbox got its metadata swapped with another image (this is a bug that happens sometimes), because it shows up as >official art, hand, <lora:artist-tsukushi akihito 2 _768px_lr-1.0e-4_epoch-000022:0.5> >Negative prompt: easynegative, (worst quality, low quality, censored:1.3), topless, nipples Which doesn't sound anything like that image and gives me pic rel.
(1.80 MB 1024x1536 catbox_cyt68i.png)

>>15807 oh sorry, that was the inpainting prompt. i forget that it does that.
>>15809 But where's the girl next door living in the haunted mansion?
>>15810 The pale-skinned gothic loli with black hair and red streaks? I forgot about her somehow.
.Anyone got resources or pointers to communities working with deepfloyd? Initial look at the results look promising and I want to help get it usable for our use cases (will require reimplementation of the inference code due to their license if we are to get large adoption for it). Implementing lora training will be another part. Been out of space for a while but finally something interesting popped up again
>>15812 >deepfloyd What the fuck is deepfloyd? Can it breathe?
>>15813 https://github.com/deep-floyd/IF Basically google image gen at home. The main tech differences is a larger text model which should in theory help create better text embeddings and super resolution (multiple upscales instead of direct 64x64 to final). The sample results makes me believe text and fingers will be handled a lot better. I had some suspicions about resolution being a key factor to character eye details but never got to the experimentation before I took a break.
>>15813 yeah that's track 2 dude
>>15814 so.. is this something that can be hacked into voldy?
>>15817 Most likely yes. Still trying to find the people working on this if they exist.
>>15817 if you can find the weights, maybe
>>15814 Oh shit, an actually brand new different base model? How exciting. Someone will probably need to do some big training run like NAI to get it working with anime though huh.
(2.28 MB 1536x1280 catbox_mmr0la.png)

Taking a break from this prompt now I swear.
>>15820 the weights aren't out yet and stability says they're trialing it to researchers first like with SDXL (which we still don't have access to the weights for lol)
>>15808 Thanks anon!
>>15822 That's how llama got leaked though right? I'm sure someone will upload it...
>>15783 Nice, did you upload it?
>>15825 not yet I got distracted with rerolling kek, still doing it only managed to get a 5 star once but it was some dude
Getting back into LoRA training, great timing I'm sure as DeepFloyd is likely gonna become the new hotness soon. I wanna make sure this seems correct: I installed torch 2.1/xformers/triton, and have xformers and AdamW8bit enabled via the CLI flags. Also copied the bitsandbytes DLLs listed in the sd-scripts readme. Training @ 768px with a batch size of 4. After a few hundred steps in it looks like the speed stabilizes to about 2.5s/it, VRAM hovering around 23GB on my 3090. Would this be considered the typical speed? I know 768px makes things take a lot longer but am curious if there's any other possible ways to improve this. Only other optimization I'm aware of that I'm not using is caching latents but all that does is save some VRAM. I'm not using the easy training scripts repo because I prefer rolling my own scripts.
>>15827 Yeah rerolling fucking sucks in this game
>>15829 Yeah, I gave up on Clara I just want a girl, that's the sucky thing about the games that aren't Honkai Impact, I just want girls not a sausage fest desu
>>15828 Batch 4 isn't needed, you can do 1024x1024 batch 3 and it might stablize between 2.2 or 1.9s/it, you really only want to stick between 2 batch or 3 batch honestly, with 2 batch you can do >1216 or 1280 res training and this is all on easyscripts for me
>>15831 I'll give lower batch sizes and higher res a try, but that speed increase doesn't sound like much of an increase to me anyway. I was using the highest batch size I could based on a suggestion the ControlNet author had made here, which was in reference to training controlnet models but I had assumed it would apply to any sort of finetuning. https://github.com/lllyasviel/ControlNet/issues/93#issuecomment-1437678784
>>15832 I'm only not sure about batch 4 because I'm not sure about the amount of steps that would work well compared to 2 or 3, otherwise it can make sense to why the controlnet author used batch 4 I would feel more comfortable with it if I had a bunch of training image to achieve a good amount of steps without too many repeats/epochs
>>15830 Eh, I like the girls in Genshin/This more than Honkai personally. Also, a good way to get who you want is to sell the rerolls with character you don't want then use that money to buy the char you do want. Just change all the info and convince yourself you rerolled it.
>>15796 Anywhere I can read more about this ddtailer? Does it make it so you don't have to always inpaint faces?
>>15836 You probably meant to reply to >>15795. Pretty sure this is the repo/PR they're referring to. https://github.com/dustysys/ddetailer/pull/28
>>15837 The Catboxes of the post I replied to also indicate they are also using ddtailer. Either way, I should have replied to both. Thanks for the link!
>>15837 >>15836 technically he replied to the right person because I posted both of these, here is my fork https://github.com/derrian-distro/ddetailer it takes the PR from the original fork and fixes up the installation stuff for it. unfortunately it still requires you to have ms build tools installed, so perhaps you might want to instead install https://github.com/Bing-su/adetailer which I think has all of the same features. it might not have the ability to put in a positive and negative prompt like my fork does but using it is really simple once it's installed. give me a moment to open up my webui so I can show you the stupidly simple interface
>>15839 These extensions essentially act as a QoL feature for inpainting right? I think using the newer Segment Anything extension which makes use of the ControlNet inpainting model might be a better choice now.
>>15840 yes, pretty much, it basically detects the face, crops it, then img2img's it, then puts it back into the image. it's faster than controlnet, and since it isn't controlnet you can actually just use it with controlnet as well. besides that, I still haven't used controlnet once... like ever. I wouldn't really compare them to be honest.
>>15841 that being said, I made my fork before the segment anything CN came out
>>15843 looks right but I can't tell you, because I haven't used it. like I said, I haven't used controlnet ever, still have no clue how it works tbh
>>15843 Yeah that's the one I was referring to. I was gonna try it out after development had matured a bit, looks like it's in a better state than it was a week ago.
>>15845 Thank you! >>15844 Thanks for the info. I'd give controlnet a shot! It's pretty self explanatory and has a lot of applications. I've mostly used to easily create a specific pose, without having to gacha + prompt perfectly. Another use for controlnet has been to recreate a background perspective consistently. I'll post an example of both these use cases combined.
>>15848 oh yeah, looks interesting. to be fair, I have seen what people are doing with it, I just haven't really opted to try it out. to be entirely honest, I hadn't felt the need to, I usually can prompt the position, give or take, I wanted, and usually background isn't that important to me... well the very specific details of the background at least. but in all seriousness, I do need to start using it, as it will likely help me improve my gens even more.
>>15833 phew finally got my Star Rail cunny Clara, gonna go get her preview images for her LORA ready and then post it
i've been making teto sing slurs all day hehe
(2.49 MB 1280x1920 catbox_866epd.png)

(2.35 MB 1280x1920 catbox_qoby4y.png)

Can I whiten out transparency pictures in paint in batches, like if I open and save them once? Or is there a nice program I can use? I got some sprites but don't want to do it for hundreds of pictures...
>>15853 Get chatgpt to write you an imagemagick script
>>15853 imagmagick magic convert -flatten *.png will flatten ALL pngs onto a white background (so keep in mind it if you have a large dataset it iterates over every image regardless of transparency)
I don't really understand how does regional prompt work. Does ADDCOL explain the prompt after or before?
(1.83 MB 1024x1536 catbox_bc686a.png)

>>15854 >>15855 thanks for the tip first try seems not bad but more testing is needed
Clara LORA released, I tried batch 4 but batch 3 ended up looking the best, I should do a test with batch 2 to 4 with my LORAs from now on, I just wish that trying to set my GPU ID to my 2nd one so I can train with both of them would work but for some weird reason it doesn't https://civitai.com/models/52688/clara-or-honkai-star-rail-lora
(1.60 MB 1024x1536 catbox_y5u2xr.png)

(1.82 MB 1024x1536 catbox_xpkdjy.png)

(1.99 MB 1024x1536 catbox_k0k1en.png)

(2.08 MB 1024x1536 catbox_6n1ilz.png)

>>15860 https://anonfiles.com/xcebT3n8zd/umineko_safetensors First try seems promising time to make BEATRIIISHEE memes
(2.20 MB 1024x1536 catbox_pktndh.png)

(1.83 MB 1024x1536 catbox_67new7.png)

(1.95 MB 1024x1536 catbox_r68uvb.png)

(1.12 MB 1024x1024 catbox_s7ygg8.png)

Made this character lora for miria marigold mackenzie for somebody. Even though I made it for somebody, I don't like keeping my lora unavailable, so I'm gonna share it here. I also wanted to share it on civitai but uh... I literally don't know how I'm gonna get around the very obviously loli character. so I think I'm just not going to upload this one to civitai. either way, enjoy: https://mega.nz/folder/CEwWDADZ#qzQPU8zj7Bf_j3sp_UeiqQ/folder/vZ50nAZS oh also, I filled out the metadata for this lora in the additional networks tab, so you can see the readme and a preview image on the lora itself if you look at it's metadata in the additional networks tab.
(2.90 MB 1280x1920 catbox_f8vprq.png)

(2.57 MB 1280x1920 catbox_xrqeof.png)

While I'm procrastinating on Roll, are there any loras people would prefer I worked on first? Right now my "up next" list is Character: Cyan (Show by Rock), Kukuri (Mahoujin Guruguru), Platelet (Cells at Work), Zero Two demon loli version (Franxx), Wiene (Danmachi) Style: krosa1028 (shennai misha), kahlua, highlow, e10, mukuton, bubukka, najucaju
>>15864 e10, why not
>>15864 >kahlua tried doing that some time ago didn't go well AI and feet don't mesh together
>>15866 >AI and feet don't mesh together the one goddamn thing i wanted the AI be decent at
(2.15 MB 1280x1920 catbox_t8biai.png)

>>15866 I've genned my fair share of feet so I think I can make it work >>15598 >>15605 >>15721 >>15791 >>15866 Just finished downloading their gallery so it'll be e10 then kahlua
>>15864 >style atte7kusa, I think its Atte793 on kemono
(3.19 MB 1280x1920 catbox_j7iowr.png)

(3.12 MB 1280x1920 catbox_vlhvpk.png)

(3.27 MB 1920x1280 catbox_tttxa2.png)

>>15869 >atte nanakusa There's one already that's decent https://anonfiles.com/x0OdTfn6z1/Atte_Nanakusa_Nai_safetensors I'll still probably do my own at some point
(1.62 MB 968x1336 01831-1028355750.png)

>>15870 Oh shit, i totally missed this then. Thanks.
It's something I guess.
>>15873 >b65 that might be your problem
Has anyone tried this pull request? https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/9256/commits Details on token merging can be found here https://github.com/dbolya/tomesd
>>15878 YOU GET: up to 2x faster gens, nearly 6 times less memory used I GET: 25% of your saturation, all of your background detail Could be useful for people with 6GB of VRAM or less but on 8GB and up it doesn't make any sense, it's like chopping a toe off so that you can run faster thanks to the rush of adrenaline
>>15874 That probably the least one since I still train the model with NAI. It's kinda rigid because I only got her LN illustration as the dataset.
(1016.54 KB 640x960 catbox_lfrmct.png)

(920.86 KB 640x960 catbox_nec8e2.png)

Quick e10 test. Unet only, no tags, no pruning, no de-jpeg-ing.
>cuda OOM when extracting with easy scripts huh? tf changed for it to not work now
>>15882 I didn't change anything about the extraction scripts, past changing the implementation to kohya's, that being said, I recently pushed an update that added back the original implementation of kohaku's version for locon extraction. so might want to update, seems like kohya's implementation might use more vram for some reason. https://github.com/derrian-distro/LoRA_Easy_Training_Scripts you can update by using the update.bat file, after that might want to also do upgrade.bat
4chan still hasn’t trained above 768? It’s over, they’re saying it looks worse because an anon that tried it out had “shitty training settings” and knowing 4chin’s standards I can’t really trust them on that
Deepfloyd doko? I have enough VRAM but I’m wondering where and how to use the model on the webui
>>15885 You aren't gonna get it working with the webui anytime soon. Clone the Huggingface space and remove all the safety check and watermark shit. https://huggingface.co/spaces/DeepFloyd/IF
>>15886 Is it possible to use it as a base for an anime fine tune in the future?
>>15887 Don't know how that would work out because they filtered out all NSFW data possible from their training dataset
>>15879 Thanks for taking a look at it anon and the analogy, perfect way to describe it
>>15888 Can’t fine tuning with hentai circumvent that?
>>15890 we'd probably have to start over from a deep danbooru finetune. hope someone's got a bunch of money to throw at it cause the non-commercial license means NAI isn't doing it for you this time.
>>15861 Nice, I'll check it out.
>>15890 >>15891 yeah we need someone with a GPU farm who is autistic and rich enough to do training with like NAI did, RIP
(2.45 MB 1280x1920 catbox_emxc16.png)

(2.22 MB 1280x1920 catbox_f93kqq.png)

(3.23 MB 1920x1280 catbox_ut4cdy.png)

(2.31 MB 1920x1280 catbox_5yudj8.png)

>>15894 Oops posted wrong pic here's a few extra
Anyone had any issues using LoRAs with Regional Prompter? Specifically it seems like once I've enabled the extension and ran a prompt with it once, I can then only ever use LoRAs on the first gen where they exist in the prompt, and then any gen afterwards no longer applies them.
(3.07 MB 1280x1920 catbox_jtwblf.png)

(2.42 MB 1280x1920 catbox_6dnd5d.png)

(3.01 MB 1280x1920 catbox_n2ry75.png)

(3.12 MB 1280x1920 catbox_3d1n09.png)

>>15894 And some more e10, sorry all Ashley
>>15773 i wonder if consumer PCs will ever be powerful enough to train a model like that? on a related note, how much money does a model like NAI cost to train?
>>15898 Many A100s. The price for just 1 of those is around $10k. Now they have H100s which are around $40k.
(2.72 MB 1920x1280 catbox_dayjhf.png)

Gluteus
>*insert autism post here* https://civitai.com/models/22530
(2.67 MB 1280x1920 catbox_aou6z7.png)

uoh
>>15897 THERE SHE IS >>15810
(1.48 MB 1024x1536 catbox_4bgc0h.png)

(1.63 MB 1024x1536 catbox_9fkpfy.png)

effort was expended to make that clothing aside work... unsurprisingly, it's kinda hard, but I made it work, for the most part.
>>15900 catbox?
oh, voldy's back. welcome back voldy, gonna give it two weeks for him to fix everything he just broke with his manic merging.
>>15909 love all ur models, thank you!
(790.62 KB 688x1032 00003-3397186742.png)

>>15894 Was wondering when someone would do e10, thanks.
>>15839 >>15840 I made a PR for adetailer to add ControlNet support and it got merged a bit ago. That should function as a lighter alternative to Segment Anything if all you care about is faces. It should have feature parity with the ddetailer fork with the benefit of extra features and actually handling anime artwork better.
>>15913 Oh nice, I'll definitely grab it when I can.
>>15909 based,thanks
>>15909 think i might need a dim16 version..
>>15909 i never thought about training on shondo's art but it does have a certain sleepy charm to it, mite be nice to mix in when making comfy non-h compositions
(1.75 MB 960x1344 catbox_90c5aa.png)

(1.26 MB 800x1120 catbox_9sp8m2.png)

(1.67 MB 960x1344 catbox_zpaovt.png)

(1.81 MB 960x1152 catbox_n6vfgi.png)

>>15918 oh yeah, this is comfy as HECK
(2.90 MB 1280x1920 catbox_4x375s.png)

>>15916 idk seems fine to me but I threw a 16 dim version in the extras folder anyway >>15910 >>15912 >>15915 I really only post here so I'm glad people are enjoying
(1.54 MB 960x1344 catbox_piesbf.png)

(1.55 MB 960x1344 catbox_elrwac.png)

(1.49 MB 1152x1152 catbox_al47z4.png)

(1.48 MB 960x1344 catbox_ms2jrp.png)

>>15920 obviously not your fault but it has a tendency of shondo-fying characters since a lot of her art is of her avatar or OCs similar to it, though this has the advantage of it working extremely well with her character loras
(2.68 MB 1280x1920 catbox_x7fajq.png)

I may need to bump Kamiko Kana up on the lora to-do list for more comf
Is the "SD is suddenly ignoring half of my tokens and I need to restart everything" bug still undocumented?
(1.57 MB 1024x1536 00010-431649871.png)

(1.55 MB 1024x1536 00042-612679689.png)

(1.96 MB 1024x1536 00043-3790407978.png)

(1.97 MB 1024x1536 00022-2966162026.png)

(1.50 MB 1024x1536 00013-3886629598.png)

(1.59 MB 1024x1536 00240-833694214.png)

(1.52 MB 1024x1536 00076-263251318.png)

(1.64 MB 1024x1536 00018-908834436.png)

(1.36 MB 1024x1280 00005-3469063663.png)

Banned from cuckchan again yay.
>>15773 hopefully we get an efficient alternative to the hugely inefficient way we train these then. holy fuck
>can't prompt pantyless pantyhose >can't prompt panties under pantyhose >it gets worse on my mix today isn't my day
>>15928 Panties under pantyhose in prompt Panties in negs. Worked for me
>>15929 nope, still getting panties over pantyhose/panties sewn to pantyhose. can't prompt them under, can't prompt the pantyhose alone
anyone have a powershell script configured for locon training arguments? don't feel like reading up on what args i need to get started with to change for my lora script but i also don't feel like installing easyscripts. no offense to easyscripts anon, your work is awesome.
>>15930 Might be a rigid lora, dunno
>>15932 it usually works tho SD is ignoring like half of my tokens today
>>15931 none taken, though it is easy to install my scripts... because I made an installer that does it for you, you just need to run it in the folder you want to install it to. though, if you don't want to do that, then you need to set the network args, which I believe is setup like network_args="conv_dim=number conv_alpha=number" but don't quote me on it, because I modify the namespace directly instead of setting it up like that.
>>15933 Never had it happen unless I had some weird setting/lora stuck I do restart my shit quite regularly because I am used to everything leaking memory
>>15879 deja vu
>>15773 has anyone done a breakdown or hypothesis on the wear and tear on the GPU from running stable diffusion 10+ hours every day? especially on a laptop, i might be taking a break here soon. my temps are consistently 5 degrees from the safe limit or whatever its called, plus idk how much its cooling down in between renders which might further wear on the solder joints
>>15937 Do you undervolt?
>>15936 meds
>>15937 As long as if your memory junction temps are below 95 your GPU will last, undervolt and play around with the mv on the core clock curve editor, you can also put your fans around 70 percent or 60 because the fans on GPUs are only super loud when they're over 70 percent imo also power limit to 80, my limit was 68 before BSOD
>>15940 Just test all of this with gens and see which one doesn't give you any crashing, or you could test it out with a modern game or one of the benchmarks
>>15940 >memory junction temps are below 95 Those are "GPU Hot Spot Temperature" right?
(2.65 MB 1280x1920 catbox_74h9s7.png)

(2.94 MB 1280x1920 catbox_6b5n4r.png)

>>15922 Looks like Kana is decently promptable on hll4
Where are places to aquires more loras? I know of https://gitgud.io/gayshit/makesomefuckingporn and https://civitai.com/ are there any more sites with large amounts of loras i can browse through or do the rest just show up on the spur of the moment in threads like this?
>>15942 Kind of but memory junction temps are focused more on the VRAM which is the thing that gets easily nuked by overheating, you can check a more accurate reading of mem junction temps with HWINFO
What command line args are you guys running for your webui?
>>15946 just --opt-sdp-attention
>>15947 no-half-vae was only meant for NAI.vae based VAEs right?
>>15948 pretty sure. i got occasional black gens before i started using kl-f8-anime2 so i'm pretty sure i'm the target for that switch and i've never had to use it.
Work on the new comfy frontend is proceeding smoothly. With some more editing options it might just become a daily driver for me
Anyone have a good script for training LORAs on the main repo of sd-scripts? I can't change the GPU I train on with easyscripts and the GUI has a maximum resolution of 1024 which I can't find the specific lines or the python file I need to change to be able to train beyond that resolution like I can with Easyscripts (but yeah changing the GPU ID with accelerate config did work with the GUI it just doesn't work with easyscripts)
>>15953 hmm, that sounds odd, oh I think I know why you couldn't change the GPU id on my scripts, my batch scripts stopped using accelerate launch because people were complaining about it. if you go into your batch file and change the command from python to accelerate launch it should use the GPU you have set. If it doesn't work, please tell me, I'll try and solve it
>>15954 ok I'll try that out, I just want to use both of my GPUs to train different LORAs, the GUI has a retarded setup so I rather keep using easyscripts
>>15954 also batch file, you mean the installer batch file for V5 or something?
>>15956 oh no I found it in the .py installer file
A lot of people say torch 2 is buggy but the only bug I have seen is caused by the more recent builds of xformers 0.18, they cause nan errors so just keep using 0.17 until that's solved
>>15957 oh no, I meant the run_popup.bat file, I don't actually know if accelerate allows you to set the GPU as an arg... I should check that out
(2.26 MB 1536x1280 catbox_l0zntn.png)

(2.32 MB 1536x1280 catbox_28zrvm.png)

(2.38 MB 1536x1280 catbox_qz783l.png)

(2.31 MB 1536x1280 catbox_tlwytc.png)

Lord have mercy on my soul....
>>15940 tfw no way to check memory junction temps on linux
>>15959 ok, so it looks like accelerate does have a specific arg for it, it's a bit clunky though, because you have to also enable multi-gpu to do it... for some reason. so in order to get it to run you need to do accelerate launch --multi_gpu=True --gpu_ids="replace_with_id" main.py --popup ... I think. so since you have two gpu's the ids are likely 0 and 1, so copy the run_popup.bat once so that you have two copies, and in it change the line python main.py --popup to accelerate launch --multi_gpu=True --gpu_ids="0" main.py --popup and in the other one change python main.py --popup with accelerate launch --multi_gpu=True --gpu_ids="1" main.py --popup then you can run them individually and it hopefully will separate out the gpus
>>15962 Didn't need --multi_gpu=true it just worked with --gpu_ids="1" thanks for the help, that's all I needed to do to train on each GPU
>>15964 great! glad it worked for you. to be honest, I couldn't really test it myself because I only have 1 gpu. I was just throwing the most likely solution out there
(4.36 MB 1320x1984 catbox_nyy4cv.png)

(2.30 MB 1024x1536 catbox_a8pgbo.png)

(2.40 MB 1024x1536 catbox_2caqam.png)

no meta on the first one because tiled controlnet washed it out so bad that I had to add saturation afterwards
>>15940 catbox?
>>15856 catbox? >>15968 not sure wtf happened there
(2.96 MB 1280x1920 catbox_urcm6n.png)

When you train a lora to give the character a chance at happiness
(2.71 MB 1280x1920 catbox_59psks.png)

(2.67 MB 1280x1920 catbox_82grl6.png)

(2.61 MB 1280x1920 catbox_tci6fi.png)

(2.65 MB 1280x1920 catbox_siah0y.png)

>>15936 i've just been in this place before
>>15943 hello fellow flea
(2.83 MB 1280x1920 catbox_tmu8lx.png)

>>15974 hewwo
>>15972 dont die i need more of these. also catbox?
>>15977 Oh no the wrong gen, shamefur dispray....
base gen was 512x768, then I used tile controlnet + adetailer to upscale + more details to 1280x1920. then threw it into img2img and used tiled diffusion with tiled vae + tile controlnet to upscale that 3x to 3840x5760. the resulting file is 21mb in size, and totally worth the almost 20 minutes of genning for this one image. I don't think i'm going to do that again, not unless it's a gen I am particularly happy with. This was more or less a test to see what I could do with multidiffusion and controlnet as I haven't used either that much, pretty damn cool. If I wanted to spend some more time on this gen, I'd like to inpaint the hands to add a bit of detail, the belly button because it just doesn't look exactly correct to me, and the panties to add more detail to them. but I'm just gonna call it good for now. https://files.catbox.moe/gi9nzp.png (file size is literally too big, 8chan throws an error for the full size image)
>>15979 based retard
(1.56 MB 1024x1536 catbox_qac3fa.png)

ghost in the shell (1995) is a lot of fun. if any bakers are looking for another idea to throw on the pile, i would love a gunbuster lora.
>>15863 way to get around it is to post only wholesome sfw preview and no loli tags/mention anywhere.
>>15982 yeah, I will probably get around to it soon enough, probably will upload it to civitai sometime today. I decided I'd rather it be more visible because I feel like I did a good job despite the lack of her in her normal outfit... or clothes in general...
Why do some style loras (like robutts and tab head) look even better when upscaled with latent? Or rather, they look even more on-model
(1.48 MB 1024x1536 catbox_azgcfv.png)

(1.77 MB 1024x1536 catbox_885ibt.png)

(1.77 MB 1024x1536 catbox_sdmllr.png)

(1.65 MB 1024x1536 catbox_vc86g6.png)

Granny(?) triple threat
>>15839 >>15913 >only works at batch size 1 huh?
>>15987 to be honest, not really that much of a detriment. you generally only want to use it once you have a gen you already want to upscale/inpaint because you can put in a prompt for it. that's what I've been doing, makes things much easier for me.
>>15988 that's a cope it works well enough with just "face, x eyes, x mouth" inpainting just the face is probably faster at that point
>Figure to try locon / lycoris training with ez scripts >Change what I think is right doesn't work >Delete and reinstall and now it says missing module huggingface.util Swear to god mang, I hate being programming illiterate.
>>15989 eh, makes a much more accurate mask than I ever could, so I'll take the cope. have fun with your version of cope >>15990 ok, so that sounds like something completely broke. I'm assuming you used the installer that I created? I want to try and help you if possible.
>>15991 Nvm I'm retarded, figured that out. Now trying to set this up to run lycoris / locon and not sure on steps, lr / unet etc to do. Do you need to change those in args or does the runner ask about those.
>>15992 all those are going to be popups, so just follow the popups. the defaults are alright, but to be honest, they were more or less set up to be used by the idiots on civitai to reduce the overall completely baked lora/locon. I suggest making a json file once you run through the popups once, because it's much easier to modify the json file than it is to modify the args directly. I can share some of my json files if you want to look at what i've made
>>15993 Yeah I'm familiar with ballparking lora with what I have I haven't touched the new stuff at all but I'm curious now since I've re-trained my shit half a dozen times now on various new discoveries.
>>15994 sure, here are catbox links to the most recent character and style lora/locon I made character: https://files.catbox.moe/ng7o8p.json style: https://files.catbox.moe/l8nhyf.json keep in mind, these don't stay the same every time for me, because I train lora by doing a test bake at 1e-3 unet, 1e-4 te for 800 steps to find expected overfitting, and to gauge how well it will learn, then rebake it with new settings after making changes to the dataset (if needed). I pretty much always use the first bake as a way to understand what the model would want to do.
>>15995 eh, should have removed the folders in the jsons, but they don't really give away anything so whatever.
>>15995 I haven't found a good reason on locon / lycoris besides the >use low dim stuff. Like what kind of LR does it like, steps etc?
>>15997 Since LoCon train on all layers, they are better for training styles, because they actually train on the layers that affect style unlike lora. take a look at the json file for the style above, I usually just do somewhere in the ballpark of 1600 steps at 5e-4 unet, 1e-4 te. but that's not in any way shape or form conclusive to what I train, because I don't train by using solely one setup. I usually do a bunch of test bakes
>>15998 oh and, I absolutely will never use dim128 lora, they all suffer from the same issues of backgrounds, hands, and eyes, being pretty shitty. If I like a style enough to download a dim128 lora, it's to fro resize it, which usually makes it somewhere in the realm of 32mb, and fixes a lot of the issues it has.
>>15999 I do 128dim on chara loras since I've felt its better, even when resizing down to 32 or 64 I notice quality drop, loss of details or namely worse hands. That was my reasoning / understanding on locon too was it was better for styles but about par for characters. Just wasn't sure if it was worth trying but I'm freed up now and don't have a character in mind I want to bake.
>>16000 yeah, when it's a character lora, if it's dim128 I just flat out don't download it, I don't even want to try and resize it to make it work at lower dims. but If I was going to resize it, I would use a ratio resize, I never do fixed resizes, ever, literally not worth it.
>>16001 but besides that, locon are technically better for everything because it trains on every layer. the reason why people don't train characters in locon is because it introduces style shift much easier. I still don't really understand why people train at such a ludicrous dim size, even more so for a character.
>>15972 what size dim do you do?
>>16002 Going through setup, what's the difference between cache latents and random crop.
>>16004 cache latents does exactly that, caches the latent representation of the images in vram, usually reduces vram spikes, not much use past that to be honest. random crop can't be used with cache latents which is why I have the mutually exclusive, but it crops images from a random corner rather then a center crop like normal, which means it should learn edges. flat out better in my opinion
>>16005 So functionally I can train locon exactly as I do loras just with lower dim? The code just handles the blocks differently? Figure I just port my settings over and see what it does, just do lower step counts. Also if I do a step count stop can I specify how many epochs or will it just do epochs based on repeats under that step count.
>>16006 it you have it do num_steps, then it will train for that many steps and stop there. it will automatically do however many epochs is required to reach that step count, this does mean that it can cut an epoch short though, so keep that in mind. when it comes to training locon though, I suggest you probably don't want to cut the number of steps from your usual, because it doesn't train any faster than lora, it just trains on more layers. if anything you might need a bit more steps because it has more layers to build weights for. when it comes to locon, I usually train at 16/8, conv 8/1, but if I'm not getting everything then I train at 32/16 conv16/8 then do a fro (or ratio if character) resize after I confirm it works. but that's only after I've confirmed that more training doesn't solve the issue without burning it.
>>15999 >>16001 Can you give a better explanation of why lower dim is better? I haven't had the issues you're describing.
>>16007 Alpha suggest was set to 1 on kohya's stuff I understood right, I've been doing 2:1 dim alpha just making sure. I'll just do my normal settings otherwise and just have lower dim and alpha 1 on the locon stuff.
>>16008 not willing to, to be honest. I've stopped caring about what others think about dim sizes, because, flat out, it's annoying to argue about something that I can just ignore by making my own! I've seen consistent improvements from dynamically resizing things down or just flat out training at lower dims, but that's not something I can prove to you. so we are gonna agree to disagree here >>16009 I only use alpha 1 if the dim is 8, otherwise I set it to half usually.
>>16010 Trying to run my json and get an error now with json.eecoder.JSONDecodeERROR: Expecting ',' delimiter: line 28. ine 346, loads return default decoder line 337 in decode obj, end = self.raw_decode line 353 in raw_decode obj, end = self.scan_once Other stuff.
>>16011 well, the error is pretty self explanatory, the json is expecting a , but there isn't one on line 28, did you edit the json file past it's creation?
>>16012 I changed the alpha and other stuff, guess I can't edit the json at all once it's been created and I just have to make a new json?
>>16013 you can, just make sure to end each line with a , that's how I usually do most of my bakes
>>16014 They all end with a comma, I only added stuff unless I'm missing something about this.
>>16015 hmm, I wonder if one of the comma's got turned into a different, identical, character, not the first time I've seen it happen
>>16016 do you think you can pass me the json file? you can scrup the file paths first, being able to see the json file means I can understand what happened easier
>>15999 i'm just a dude that wants to learn, not argue
>>16018 I guess that's fair, I've just become pretty jaded about it because, well, maintaining the scripts, you get to hear a lot about how shit you and your lora training methods are. it's one of the big reasons why I just don't fight people on it and let them do what they want. I can just opt to not use it, or resize it if I really want to use it.
>>16017 The problem fixed itself by just making a new json, but i'm having issues with batch training. How do I need to set this up to actually do multiple loras instead of just amalgamating one big meme lora.
>>16020 you need to put them all into seperate folders, and train them individually, one at a time, you can do this by creating a bunch of json files, 1 per concept, and putting them all into the same folder. when the popups asks you if you want to do multiple trainings, then you can just give it that folder and it will train them back to back. this is just unfortunately the best option I was able to give you because it's tough to keep track of the variables for any number of trainings in the queue.
(1.25 MB 1280x896 00027-220175003.png)

>>16021 I did that, copy and pasted the json, changed each one to a separate concept folder and it's returning me the same error I had earlier. It's like editing the json just bricks it somehow.
There was only one great thing I found off of 4chan recently, it was the Devolver-chan LORA, only 8 images but the way they made it was 10 repeats and 10 epochs, I'm surprised it isn't fried with those repeats but this will be a good LORA to study https://mega.nz/folder/ctl1FYoK#BYlpywutnH4psbEcxsbLLg/folder/R58yTYLY
The eye glow on hair from the Fiz Rot LoRA can be an interesting effect, though very gacha.
>>16022 Love this. Fat bastard's anatomy is a bit incoherent but the scene and her face are just great
>>16002 >why people train at such a ludicrous dim size it just werks also big and varied enough dataset there is basically no embedded style
I had some fun, I may improve it further. Also this cunny tag team is making me ToT.
>>16023 It must be your editor then, because I modify my json all the time.
>>16028 But you can do that at a lower dim and probably get better results... but then again I don't really care I guess. People can do what they want, I'll just put actual effort into my lora.
>>16031 what is the point of this post lmao
>>16029 CATBOX PLEASE
All LORAs are shit, they lost their uniqueness now so there's really no need to give a shit about dims or alphas anymore. We need to find the next cool thing for AI image tech
>>16032 Just calling people who train at dim128 lazy, that's all. >>16034 Won't change the fact that I don't use dim128 lora because they are pretty consistently worse than using a resized version. Also... the "next thing" for ai art pretty much can only be achieved with an architecture change.
(2.17 MB 1536x1280 catbox_3ii5k6.png)

(2.31 MB 1536x1280 catbox_2jjdoq.png)

(2.17 MB 1536x1280 catbox_uvilhr.png)

(2.03 MB 1536x1280 catbox_h0dsj1.png)

>>16033 https://files.catbox.moe/0fkd7h.png Boy do I love this prompt but I should work on some of my other ideas, like breasts on glass with several characters and the yuri river prompt.
>>15883 >tried extracting again, OOM again AAAAAAAA it takes forever on CPU
>>16037 how large are the models you are trying to load? if they are both 7gb for example, you are trying to load 14gb of models into vram
>>16038 two pruned 2gb models working on 6gb vram (again, worked before)
>>16035 what do you mean resized?
>>16039 hmm, that is odd. I know I didn't make any changes to the code I wrote, so it might just be something kohya changed. >>16040 using the lora resize scripts that I have made which use kohyas resize script for lora and locon. https://github.com/derrian-distro/LoRA_Easy_Training_Scripts/blob/main/lora_resize.py it allows you to dynamically resize lora, and flat out, a dim128 lora has a ton of empty space in it, which can easily be taken out without changing anything about the model except for it being like half the size a minimum. I usually go further than that and do a sv_fro dynamic resize at 0.85 max dim 16, and usually don't get much of a reduction in quality if any. if your goal is to not touch the model quality and just reduce size, a sv_fro at 0.99 max dim 128 will work for that, and should cut the file size in half give or take.
>>16040 why not just start out with a 64 DIM model to begin with?
>>16042 8 dim 1 alpha
>>16030 It opens with code writer whatever that is, think it's just a default windows thing?
>>16044 I don't think "code writer" is a default windows thing, I personally use notepad++ to open up my json files
>SAG extension so shit it changes main files, preventing you from gitpull yeah, no thanks
>>16050 Learn how to git merge
>>16051 >you NEED to change how updating works because dev of abandonware extension is too retarded to code properly no no extension ever needed anything more than a git pull
is it safe to git pull? i'm on an old commit from like february, before the gradio updates/changes i've seen people complain about
is SDXL "just" a new model or will it require some fuckery to get it to work once it's released and monstrous amounts of vram?
>>16054 SD is only 859 million parameters, it's safe to assume anything newer will be bigger
(1.53 MB 1024x1536 catbox_u3ypkb.png)

hey onii-chan, you come here often?
(3.18 MB 1280x1920 catbox_36v39w.png)

(3.30 MB 1280x1920 catbox_17zwur.png)

Thanks to the og anon who made the Nanahira lora. Here's a new version using only art with her current design, getting bebe to show up is practically RNG though. https://mega.nz/folder/2FlygZLC#ZsBLBv4Py3zLWHOejvF2EA/folder/KBtDzY4S
(3.87 MB 1280x1920 catbox_8s5sf1.png)

>>16057 UUOOHHH😭😭💢💢
Controlnet post about more batch being better and not going lower is bs, literally anything above batch 3 is only better when you have way too many images in your training dataset
>>16057 og author here. recently made a new one too, just testing which one is better but I'll compare yours with mine to see if I keep it under wraps or not
>>16059 wonder how many times people are gonna have to learn that beyond a certain minimum, higher quality data is more important than more data before it sticks
>>16060 I can share my dataset including images I didn't use if you'd like. I didn't end up using a lot but everything is cleaned up and scaled (max 2048px)
>>16063 yeah sure, I want to see how many images you ended up with
>>16064 64 images total, 43 in my final version. Only manual tagging was to add nanahira and bebe tags, though bebe didn't really do much. https://anonfiles.com/WejdQ3o6zd/training_image_folder_7z
>>16063 makes sense why you would get the bebe zipper more, I kept with my 22 image dataset
>>16065 curious though what is your process for upscaling training images?
>>16059 >>16062 Did anyone ever test batch 1 vs 2 vs 3? I don't mind training slower if it's gonna make the lora ""better""
>>16068 I can test it with the Nanahira rebake I'm doing honestly I still like batch 2 the best because I can go as high as 1280 with resolution, batch 1 I'm still not sure about the limits desu
>>16067 I don't upscale, I just downscale anything that's too big. I also usually I enable bucket_no_upscale though I'm still on the fence about if it's better or not (though having it enabled does make training faster). It was more of a problem at lower training resolutions where you would get bad aliasing so as a rule of thumb I just downscale to max 2x the width/height of the training resolution. I don't think it's actually worth upscaling training images because you won't get any more detail that's not just artifacts from the upscaler. SD does a good job at filling in the details anyway which you can see pretty well with the figure loras. It keeps the ISO noise and plastic texture at higher resolutions even though the training images are much smaller.
(1.10 MB 640x960 catbox_5vlbk6.png)

It tried
>>15389 What's the model you are using for the first 2 pics?
(513.73 KB 512x768 pyy4yb.png)

(411.09 KB 512x768 4a7kti.png)

>>16063 Your Nanahira looks more Loli kek, I'm surprised your LORA wasn't fried with 1e-3, batch 1 maybe? I tried it out with batch 3 and got the fried green pixel mess
>>16073 Could be because I only use batch 1, are you using the same dim size? The only other setting I've found to work with 1e-3 is 16 dim, 8 alpha. I've kind of been stuck on these settings for a while because I haven't been able to get anything better yet at 16 or 32 dim. On the plus side it's a lot easier on my HDD space. Also I'm doing one more batch with a few more images and a few cropped images of bebe with some additional manual tags added, specifically rabbit, holding animal, animal hug, and animal on head. If I can't get consistent bebe after that then I give up.
>>16075 I've stuck with 64 dim, it's been awhile since I used lower dims but yeah your post makes sense that lower learning rate equations wouldn't fry at lower dims like 16. I just mostly want to focus on genning Nanahira and making sure she can change her looks easily, good luck with Bebe though I have to test this LORA out now
(2.50 MB 1280x1920 catbox_gl8a18.png)

(759.86 KB 640x960 catbox_fh82mu.png)

Still no luck. >>16076 I used 64 dim for a while with locon and lion and got some good loras out of it at 5e-5, namely musouduki and tsukushi akihito
(480.65 KB 512x768 00343-1610153410.png)

(429.97 KB 512x768 00362-2965841007.png)

>>16078 jesus christ how horrifying
since I noticed here and on 4chan we have anons running 2 3090s or 2 4090s have any of you anons decided to run multi-gpus setup? Doesn't have to be expensive cards, multi gpu setups might be more viable now also due to language models working with SLI/NVlink not sure about AMD crossfire though
>>16080 i'm building a 2080 ti / p40 setup for language models
>>16081 Not bad, P40s are not that bad on the used market and the 2080ti NVlinked was always cool to see back then
>>16082 should be able to run any ~30b model without problems and i've got hopes for sparsification taking me higher than that
>>16077 I tested out -e3 with lower dims like 16 and yeah it doesn't fry even with batch 3, that's weird honestly I didn't know higher dims could cause such an effect with lower exponents. I'll see how 32 dim looks because it lets me train 1080x at batch 3 faster compared to 64 dim.
(879.04 KB 768x1152 00118-2412711289.png)

(953.92 KB 768x1152 00093-4006510786.png)

(816.09 KB 768x1152 00109-2096436873.png)

Baked some new Loras today, Nowi and Faa from Fire Emblem and Serena from Yugioh Arc V. https://mega.nz/folder/pZ9nBJ5Y#cmpb-DWVSWqqXg3fi9zX2Q
(1.92 MB 1024x1536 catbox_5wjow4.png)

>>16086 Nice
>>16062 thing is, what is the optimal amount of images? What is a good image and what is a bad one? Is a CG better than a sprite? This is what keeps me up at night. >>16068 the few tests I did dont really seem to have that much of a difference, I liked how batch1 handled things in comparison to 2 and 3 and I think it involves less autism with adjusting LRs to batch sizes. >>16070 >upscaling I remember someone saying super small images like 800x600 are better upscaled even if ur training on 768, have I simply been clowned?
>>16084 the other day I saw someone run lower dims on 5e-4, tried it out on 32-16 and it seemed better. Today I saw >>16075 and tried out 1e-3. Well, maybe I should have let it bake longer but it doesnt seem right. I also have to question whether using 4e-4TE wouldnt be better, across all my tests I've at least learned that it needs to be lower. In conclusion I think the peak lies somewhere around messing with high LRs and alpha values that will let you use them, have you tried running this on higher dimensions?
>>16089 I'm always running higher dimensions now and honestly not sure if learning rates would make as huge as a difference compared to what batch and dim/alpha you are using but I have just been using alpha = dim recently since I noticed it made my gens look "slightly better"
Apparently the new pull Voldy pulled out is making people switch over to torch 2.0, not sure if this is true but I heard the tagger WD extension is broken right now with the latest pull so hold off for a couple of weeks
>>16092 anon we've been on torch 2.0 for more than a month now tf
>>16094 hey if you enjoy slower it/s, more power to you.
>>16092 works for me
>>16095 torch 2.0 has vastly increased vram usage, I'd revert if not for extensions
(3.39 MB 2560x1118 catbox_k168m6.png)

Is there a good way to save every X steps instead of every epoch? I haven't played with resume yet but I also don't want to meticulously set up multiple json files
>>16089 was there even a huge difference between torch 2.0 and 2.1 in terms of speed?
>>16095 >>16103 wrong reply
(2.00 MB 1440x1440 catbox_dcw93o.png)

Not super happy with it but here's allcharacters18 (fisticuffsclub) https://mega.nz/folder/uZkBHZqS#qxCUIezviHmIT3OLCaVocg
Still trying to find out what the best learning rate is and at this rate I just gave up and did the exponent equal to whatever my batch is, it's just a slight schizo difference for adamw compared to lion but the text encoder is the only thing making me super schizo wondering if 5e-5 should just be left alone or changed
There's a lot of new models being uploaded, I have to make Based66 now but I'm realizing a lot of these people are using cool forms of MBW but a lot of new mixes are having LORAs merged into them, Counterfeit 3.0 and Aurora mix. I'll have to look at the MBW research to see how I can work with newer models I like while avoiding too many finetunes or LORA merging because I don't want another situation like Based65 however Aurora's posts are showing some LORAs details are showing https://civitai.com/models/40199/aurora I have no plans of mixing style LORAs into Based66 but I'm impressed some LORAs are working with this mix despite that crazy conspiracy chart tier mixing recipe
(3.05 MB 1280x1920 catbox_ndj8ox.png)

(3.15 MB 1280x1920 catbox_lj4g9g.png)

(2.99 MB 1280x1920 catbox_rse8g7.png)

>>16109 Are there any models your would recommend right now? I've been happy with AOM2 and Based64 for a while now but I think it's been long enough to try something new.
>>16110 I haven't been posting a lot of LORAs or outputs for pixiv so I just stuck with Based64/65, with new models I'm not sure because I haven't tested them out but I've always liked the earlier Defmix models (seems like the owner's last model was 2.0 which put counterfeit 2.0 into the mix so putting it into my mixes would affect LORA compatibility)
Just when I thought of mixing supermerger isn't loading my models for mixing. Ugh I'll start working on Based66 soon because I want to use a different model for what I post online
>>16097 yeah but if you have enough vram then it doesn't really matter. disclaimer: i have 24gb
(418.31 KB 828x842 1682900966942427.png)

>>16086 >Nowi
Are there better "multi purpose models? Or does Based64 still reign supreme?
>>16114 I might be retarded, but I can't figure out who this image is mocking
>>16116 the one underneath is from twittertards who try to look morally superior (compared to the 5 IQ retards who just shout pedo left and right) when they explain why you should be arrested for liking cunny, the one on top is self-explanatory
>>16115 Earlier AOM2 merges, there's many but I'm pretty sure all of them work well with multipurpose uses
>>16115 7th layer v2a never let me down
Why is it so hard to force colors on clothes sometimes?
>check "main" thread >see some of the most deformed civitai-level shit >close tab
>povImminentPenetration_ipv where can i get this lora maid anon?
>>16124 It used to be up on civitai but it seems like it got nuked: https://anonfiles.com/EfUel2p0z9/povImminentPenetration_ipv1_safetensors Dropping penis on face as well cuz that shit also got nuked: https://anonfiles.com/D8V4lbpbzf/penisOnFace_penisOnFaceV2_safetensors
>keep getting extra hands on chest instead WHY
>>16126 you clearly don't want help, because you didn't post the whole prompt and the result. so posting this is just looking for attention, right?
>>16127 take your snarky remark and stick it up your ass, faggot
>>16121 >coming home tired after a long day of work, you notice the lights are out in your modest Japari Park bungalow >"hmm, that's strange, it's not like shoebill to go to bed this early" you think to yourself >after fumbling with the keys you open the door and groggily step inside >"honey I'm home!" you call out >no reply... >you call out again, louder this time >suddenly, the silence is shattered by the sound of gunfire and screaming coming from the master bedroom >"Shoebill!" you cry as you dash down the hall, pausing only to draw your Japari Park Ranger standard issue anti-Cerulean Glock >kicking open the door to your bedroom, you flip on the lights, ready to confront the intruder >instead, you find the source of the noise >your wife, Shoebill, is lying on the bed you share >perplexed for a moment by the gun in your hands, she then relaxes, looking up to meet your gaze >that cold stare, which would otherwise frighten a stranger, to your familiar eyes betrays a fiercely loving expression >dressed in naught but a négligée, she lifts a hand to beckon you from its resting place on her toned stomach >"Oh honey, what will the neighbors think?" you remark as you holster your Japari Park Ranger standard issue anti-Cerulean Glock >That night the whole neighborhood is kept awake by the sounds of automatic gunfire and screaming as you lovingly rail your wife until dawn https://youtu.be/kDMHHw8JqLE
>>16129 >>That night the whole neighborhood is kept awake by the sounds of automatic gunfire and screaming as you lovingly rail your wife until dawn Okay, I laughed.
>>16131 nothing is cheating
>>16131 pay the fine
>>16134 you will never have my 350 bucks
>>16133 >"nose blush"
>>16136 ¯\_(ツ)_/¯
>>16137 a dangerous upscale tag
>>16140 do you have to make the same post on 2 threads? I hope you posted it on twitter and pixiv too
Supermerger with the MBW presets is so cool, now I can see the difference of each preset without having to save the model. Based66 might come out next week I just have to be careful of newer models that are LORA merged not sure if that will effect LORA compatibility too much but Aurora mix seems to do fine with LORAs so I'll see what happens
>>16141 i'm pretty sure a lot of this board no longer checks the /h/ threads regularly
>>16141 What >>16143 said. I check it once every few weeks and just post stuff there too whenever the thread isn't 90% grannyposters from /g/
>>16144 >>16141 >I hope you posted it on twitter and pixiv too Glad it bothers you enough to make a cute passive-aggressive remark about it, I'll keep doing it just to annoy you.
>>16141 I'm going to fucking rail you until dawn
>>16142 Yeah LORA compat is definitely high priority if its something that the entire community uses and not just a model that can prompt 1 or 2 things, great stuff btw.
Despite using OUTD with my shitty Shoebill LoRA the takorin/csrb/fizrot mix I wanted to showcase didn't come through that well so here's more Cirno instead.
>>16148 Very cute
>>16148 Despite mixing three style LoRAs the takorin one still prevents me from prompting Parsee.
>>16142 That Aurora kinda struggles when you try to mix sexo concept lora and character lora. It's just forget the male existence.
(3.52 MB 1280x1920 catbox_2d51mu.png)

(2.82 MB 1280x1920 catbox_4zdaaw.png)

Kukuri a cute
>>16151 Really? If any of you can provide any sex concept LORA with males so I can test it out with based66 that would be nice
>>16145 what part of that post implies it bothers me lol im just making fun of you dumbass >>16143 doesnt really seem that way, dunno why would people stop anyhow
>>16155 I'll test it out but I'll also release the first version of Based66 I have here so anyone with more LORAs or LORAs collectors from here can test it out themselves so I can find out how good the compatibility is
>>16156 I hope you also put it on hugginface so I can easily fetch it as collabfag.
>>16157 Yeah I'll do that since I upload all versions of my mixes onto huggingface compared to Civitai for archiving purposes
Only messed up with one gen with the 1 twintail like Based65 got way too many times, I'll go back to the drawing broad
Which LoRA Block Weight preset can help with removing baked-in clothes from a STYLE LoRA?
>>16160 consider testing on more than one lora - have you at least checked how it behaves with heavy style models to know if its even good for testing?
>>16162 Any suggestion for heavy style LORAs? I only collect/make character LORAs
>>16164 >heavy style loras just any style lora? if you meant heavy style models to compare then I dont test too many models but I think Silicon fits the bill
>>16164 afrobull, zankuro, houtengeki, merunyaa, kairunoburogu, pumpkinspice, squeezable, and muk are the ones that I can easily think of. There's definitely more
>>16165 you can try the saya no uta lora that imitates old game artstyle, I think it was on hdg repo somewhere
>>16167 meant for >>16164
>>16148 do you mind sharing the mix weights?
does anyone have "Based64mix-V3_ema-560000_fp16"? I've got based64mix-v3 already, is there any difference between the two models?
>>16170 Based on the filename, it's just based64 with vae ema56000 baked. You can just download the vae separately instead.
(3.25 MB 1920x1213 catbox_ospao7.png)

>>16170 >>16171 Can confirm, I run every model I download through model toolkit with "Fix broken CLIP position IDs" and "Enable Autopruning" enabled as well as changing out the vae. >>16164 I would say wagashi v2 or any style lora which works well on base NAI, off the top of my head the mizumizuni and og love live loras
>>16166 >>16172 Are all of these on the gitgud repo?
(4.75 MB 1920x2505 catbox_jr826k.png)

(5.10 MB 1920x2505 catbox_8lqv2j.png)

(5.01 MB 1920x2505 catbox_z8a1ab.png)

>>16173 wagashiv2: https://mega.nz/folder/dfBQDDBJ#3RLMrU3gZmO6uj167o-YZg/file/UOgVSTCJ The mizumizuni one was taken off civitai but there's anew uncensored one that's more compatible so it might not be as good for comparisons https://civitai.com/models/47123/deepthroat-blowjob-pov-fellatio-mizumizuni-or-sex-act-lora-188 Also the love live one isn't quite what I remembered but it's still good: https://archived.moe/h/thread/7076593/#7077215
(2.71 MB 2160x1222 catbox_hqfs3h.png)

(5.47 MB 2400x2624 catbox_dnnf3z.png)

>>16174 I still have 2 versions of the one taken off civitai that I reupped if anyone still wants them >>13325 >>13328 https://files.catbox.moe/9nugyz.safetensors (newest before deletion) https://files.catbox.moe/afophj.safetensors (original)
>>16174 >>16175 thanks I'll test them out alone and with character LORAs
anyone experience an issue where sometimes it/s for lora training is really slow compared to normal (normally i'm training loras at like 1.85it/s, but when the issue happens it might not even reach 1it/s or if it does, very slowly)? nothing indicates something is using vram, but restarting the pc fixes the issue.
(1.54 MB 1296x1296 catbox_nem9c0.png)

>>16171 >>16172 anyone mind uploading the model? even with the vae it seems pretty hard to replicate >>16106
(827.54 KB 768x1152 00022-925576957.png)

big
(3.16 MB 1280x1920 catbox_t5c1vx.png)

(2.20 MB 1280x1920 catbox_rxbnde.png)

(3.49 MB 1280x1920 catbox_3heavs.png)

(1.70 MB 1024x1536 catbox_xq1o5v.png)

I decided to take a break from working on the UI to make a style lora. I think it turned out pretty good, struggles a bit with sex, and his unique way of drawing pussy has pretty much solely confused the model, but when you prompt cleft of venus it works really well. Works well with sex lora, so not a huge issue. I like the whole oil-painty style it gives off. the artist is utatanelolicoco links: https://civitai.com/models/58063/oil-paint-anime-style-locon-or-utatane mega: https://mega.nz/folder/CEwWDADZ#qzQPU8zj7Bf_j3sp_UeiqQ/folder/Kdoy0LJI
>>16174 >but there's anew uncensored one that's more compatible Where?
>>16180 hi shank
>>16182 >struggles a bit with sex me too thanks :(
(1.17 MB 878x1280 00016-1292444679_cleanup.png)

(1.21 MB 1024x1280 00014-3243275448.png)

(1.30 MB 1024x1280 00013-3307127020ed.png)

bless the gomennasai lora
does anyone got the zackary911, syuro, hanuvo, tabhead loras? cant seem to find em anywhere
>>16188 >zackary911, syuro anon.....................................................
>>16191 ah. Thought you linked the dead one, whoops
>>16188 tab head I have in my mega drive and on my civitai page: https://civitai.com/models/18748/tab-head-locon https://mega.nz/folder/CEwWDADZ#qzQPU8zj7Bf_j3sp_UeiqQ/folder/zYo33AbD I personally believe that the locon version I cooked is much better than the dim128 version I cooked back when lora was new. (and being only 14mb)
Based66 is done, I managed to fix outputs for one seed that was bothering me when mixing a character LORA and style LORA. Got both details rights, I'm making the preview images for the Civitai version but I'll upload the unpruned version onto hugging face and post the link here for anons here to test it out. I'll upload the pruned version onto the Civitai upload and huggingface too later.
Chinks came out with something crazy
>>16187 catbox plz sir?
(2.06 MB 1024x1536 00181-2841540907.png)

flat
(1.87 MB 1024x1536 00233-3930750439.png)

>>16201 not flat >>16196 Thank you for working on this! Hoping this will be a nice improvement over Based65.
(1.62 MB 1024x1536 00349-3718971226.png)

>>16203 Very important to get the right measurements for a proper fit.
sex with cirno so much sex with cirno ungodly amounts of sex with cirno
Based66 unpruned, good luck colab bros (I'll upload the pruned version on civitai and huggingface once I have the preview images ready) https://huggingface.co/AnonymousM/Based-mixes/blob/main/Based66-V1.safetensors
>>16208 Yeyyyy
>>16210 not directed towards you but god i am SO fucking tired of the AI giving me undressing/taking jacket off/clothing hanging off of the arms despite having everything related in the negatives
>>16212 Undressing in general is kinda wonky for the AI.
>>16205 I just regenned this image from the one with metadata on aibooru, removed the cirno LoRA, and got an identical result, lol.
>>16214 Oh, now I see why too. Did nobody fucking check the metadata? It's literally just a 2gb file of empty weights.
>>16214 >>16215 I was wondering why it seemed to do absolutely fucking nothing, I thought it'd help me correct (uoh) Takorin's sometimes wonky/incorrect colors and it did nothing even when turned up to 2.
>>16215 >2GB lora bro...
>>16216 >>16214 >>16215 THIS happened by complete accident too, just tested it without the 2gb LoRA. I'm wondering if the guy who trained it was just fucking with us.
(169.53 KB 1036x898 msedge_9G3lqkbTqA.png)

he got us
(5.55 MB 2560x2424 catbox_lly17q.png)

(4.79 MB 1760x3133 catbox_5ztiwa.png)

(1.75 MB 1280x1213 catbox_85e5ix.png)

(4.50 MB 1760x3133 catbox_qbi2ao.png)

I might use based66 occasionally but probably gonna be sticking with based64. It's just too good.
>>16220 oof that one Nahida seed that didn't get her ear, I'll probably work on a better version soon it just takes such a long time to bake a mix and see what seed should be focused on along with other seeds to see where the mix could do better.
I can't stop using Based64 Final and Based65. Can you guys recommend some similar models or any of your favs to try? Don't really want prompt for sex, just casual nudity or erotica or fully clothed (lol).
>>16221 did you mix the vtuber one into it?
>>16223 yeah the last epoch for HLL4 epoch 8 for p3 I think
>>16208 done like 20ish prompts on it so far and it honestly just feels like a worse version of aom2 or anyhentai so far all around, what exactly went into the making of this?
>>16225 hll4 nutmeg aom2 defmix just whatever I could use that avoided too many finetuned models to avoid the problems B65 had. I'll probably work on another version, you can share a zip of your outputs and whatever issues you had so I know what to fix
>>16226 isn't defmix fucked? or was that refslave?
>>16227 defmix 1.1 and beyond, 1.0 and all the colors didn't have finetunes (Counterfeit) Refslave was the finetune
>>16228 idk chief, even 1.0 seems fucky
>>16229 red was what I used for based64, it was good for NSFW outputs not sure if it would work the same way as it did back then given hll4 is a new factor.
>>16230 but if it doesn't work well in this case I'll look over the mix and see if I should keep it in and take it out, I would like to take a look at those outputs to see the issues and use the seeds for x/y grid for the 2nd version of the mix
>>16226 doubt its worth sharing outputs since its just basic crap with diff poses for the sake of seeing style/anatomy of the model, but tl;dr of what i seen on my end: model seens better style wise imo, felt less 3d and more animey/drawn for the lack of better terms, but anatomy feels pretty scuffed compared to b64 esp hands(yeah yeah i know "hands are hands yada yada"), faces felt less detailed imo, lighting seens to have taken a small hit but its prolly cause of the style change but yeah prolly just a matter of playing with weights/blocks of things you are merging, from the looks of faces in particular i feel like maybe defmix is coming out too strong on it? none the less thanks for your hard work as usual anon
>>16218 2gb of "hey you fuckin retards you don't need a lora for characters the model ALREADY KNOWS" what a hero
>>16233 nice bait but not today anon
I'm not lazy, I'm efficient.
(2.04 MB 1024x1536 catbox_2z2rwj.png)

>>16208 will play around thanks anon
>>16236 post any outputs that have problems and share them with me, I'll look over it and see if V2 can fix issues with anatomy and hands
how to improve your aibooru experience in one easy step
>>16238 There's only one kind of person I hate more than faggots who upload shit to aibooru just to shill their grift. It's faggots who upload shit to aibooru just to shill their grift and don't provide the metadata for it for whatever reason (deleted on purpose, touched up in PS, jpg/lowres to encourage you to sub to their grift, etc) And jews and niggers but that's not the point.
>>16239 this, holy fuck >if you want metadata subscribe to my fanbox/patreon! Fuck you
(2.19 MB 1024x1536 catbox_p2ztby.png)

(1.99 MB 1024x1536 catbox_zj1qi6.png)

>>16237 Did a little bit of testing with one of my river prompts, I will throw some more prompts at it later.
(2.21 MB 1024x1536 catbox_b4r4t6.png)

>>16237 will do but for now it seems fine for my standards without lora
>>16240 I blame the site as well, either provide the metadata or the image won't be approved. Then again, someone on the staff thinks that shoving the SLIGHTEST amount of cleavage or bare arms is enough for a "sensitive" rating rather than a "general" one.
>>16242 As long as it's fine, I can fix issues later with V2 or B77 depending on what direction I go but yeah this was the first time I really focused on MBW with supermerger's x/y plot but if you guys have any bad seeds share the catbox and I'll include it for testing
>>16244 kinda fascinating to look at the differences
>>16245 b66 is getting close to that pastel but sharper look I wanted with the lines, I might make V2 myself aside from any anatomy issues due to specific models or MBW presets to get an even more sharper look, 7th anime C was the mix I was focusing on along with expline mix to get what I wanted.
I'm keeping V1 of B66 on huggingface only, I'll upload the pruned version there but Civitai gets V2 which fixes some issues and takes out some models that might affect the mix's performance
I can't make sexo ones since I'm on cagie, but comparing it to other based version. It seems like it didn't capture the loli essence without prompting it. Also, it added some detail that you didn't ask(random panties string) like based65(bag even though you didn't prompt it, probably loras layer). It can be good or worse depends on gacha and loras quality. >>16248 Thanks anon, waiting for v2
>>16249 Forgot to add, the artsyle probably the best of the bunch. I like how it nails the contrast
>>16250 do you have the LORA and image for that B66 output? This could help with small details and anatomy while I'm working on V2
>>16252 thanks, the only big change I'm making with V2 is taking out defmix because I found out even the colored versions are using the finetuned Counterfeit model and I want to make sure the vtuber outputs from HLL4 still work well
>>16253 I remember some good loras, but because it was >trained on counterfeit. It just became wonky when you try it on model other than counterfeit itself. You probably can take out the schizo blocks but probably more trouble than its worth.
>>16254 doing custom blocks myself would make mixing longer, I'll probably do it for B77, I'm just trying presets and x/y grids that will help the most. Not sure what B77 will do differently because sadly almost every model is starting to do LORA merging and Counterfeit 3.0 adding which is just going to make mixing and LORA testing harder. By that time if those are the only models left I'll have to start doing schizo block mixing to see if it will fix issues
(1.75 MB 1024x1536 00159-319039258.png)

(1.74 MB 1024x1536 00003-2408185597.png)

>>16208 >>16248 Thank you for the release! I feel like I'm getting fewer anatomical errors than when using Based65, but still somewhat more than Based64-V3. Here are some catboxes for testing reference: https://litter.catbox.moe/8584jz.png https://litter.catbox.moe/2b4mtv.png LoRA models: https://github.com/AUTOMATIC1111/stable-diffusion-webui/discussions/9084 https://civitai.com/models/13465/taihou-all-skins-azur-lane Looking forward to V2! >>16241 Gorgeous flatties!
>>16256 Thanks this is very helpful, I'll be sure V2 fixes these issues.
(1.74 MB 1024x1536 00013-2473102803.png)

>>16257 Glad to be of assistance!
(3.90 MB 1536x2304 catbox_q409xp.png)

(4.01 MB 1536x2304 catbox_4n9apn.png)

>>16253 out of curiosity i ended doing my own block merge of the models you listed except for defmix then added 1 personal model of my own and ended getting good results of it, pretty sure defmix is causing most of the problems in your v1 prompt stolen from river anon up there minus the loras
It's kinda shy a little more on nailing lewd concept compared to other 2. Based65 adding mole on her tiddies, kinda lewd meanwhile based66 adding additional blue tint and based64 just makes zurisussy which is sexo. I feel based66 reaching anything domain instead, which isn't really all that bad.
>>16259 Yeah realized counterfeit is a really dicey model when it comes to mixing and all versions of defmix was using it too. 3.0 is going to make a lot of model mixes uploaded to normie sites annoying with LORAs if they don't know how to do the right block merges to avoid the LORA troubles it has.
(2.94 MB 1280x1920 catbox_e8k9z3.png)

(3.09 MB 1280x1920 catbox_zgrzti.png)

(2.68 MB 1280x1920 catbox_r6vvpn.png)

(2.81 MB 1280x1920 catbox_h2mbpj.png)

Finally got around to releasing the Arona + Plana model I was working on for a while now (tagging was hell because I put them both in the same model, and opted to do a "combo" folder) but it seems to work pretty well. it has it's quirks, but it's flexible for the most part and can do both characters perfectly most of the time. I included a readme in the mega drive, civitai, and the model metadata itself. I know some people were asking me to make Plana back when I made Shun + Shunny, but I only really had time every so often to work on it between working on the UI stuff and other projects. hopefully they are still around. anyways, links for people to download: https://mega.nz/folder/CEwWDADZ#qzQPU8zj7Bf_j3sp_UeiqQ/folder/qUpFmABI civit: https://civitai.com/models/60885/arona-and-plana-or-blue-archive-or-character-locon hope you enjoy using it, I know I enjoyed making it (despite the tagging nightmare)


Forms
Delete
Report
Quick Reply