/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.

(1.86 MB 1024x1536 00008-2178941320.png)

/hdg/ #14 Anonymous 05/08/2023 (Mon) 09:59:19 No. 16264
Previous thread >>14999
(1.64 MB 1024x1536 00233-3824712532.png)

flat
If anyone wants to test it out before I upload it to Civitai https://huggingface.co/AnonymousM/Based-mixes/blob/main/Based66-V2-Pruned.safetensors
>>16267 Thanks anon, it's too faithful to the loras and kinda forgot the prompt. Which you probably can just put more weight to the prompt or play around the CFG I guess. Haven't tested that much but overall better already.
>>16268 >too faithful to the LORA first time I've heard of that in awhile given all the new finetuned models and LORA merging people are doing with recent mixes kek
I don't know if anyone here cares but I've been posting AI Gura voice clips on /vt/ in the gura thread. Originally just did I cover of her singing Mousou Express but it kind of escalated to say the least so fair warning Mousou express: https://voca.ro/19Xi68GCmGdX Voice clips: https://voca.ro/156vLyKMWbHq https://voca.ro/1nCUYRdn50Hm
>>16270 Was this using Bark or something? I'm still only keeping my eye on Art AI tech given how easy it is to run
>>16271 It's RVC, basically a better version of sovits or diff sfc. I'll copy what I posted over there This RVC model that was just uploaded today https://huggingface.co/QuickWick/Music-AI-Voices/tree/main/Gawr%20Gura%20(RVC)%20126%20Epoch It seems like it was trained on her ASMR since it can only really do her whispering which is why I did Musou Express. If you actually want to use it it's pretty easy, download RVC-beta.7z from https://huggingface.co/lj1995/VoiceConversionWebUI/tree/main It's not TTS so yes I did have to use my voice. Though I guess you could inference from TTS
(1.80 MB 1024x1536 00039-1590909576.png)

(1.62 MB 1024x1536 00006-623350991.png)

(1.68 MB 1024x1536 00008-3401207129.png)

(1.78 MB 943x1536 00008-3802945083ed.png)

>>16199 https://files.catbox.moe/588f5v.png https://files.catbox.moe/pznss3.png https://files.catbox.moe/w3z429.png Excuse the slow reply. Basically used other anon's "Remi in negligee" prompt
>>16272 >had to use my own voice >these moans im gonna vomit
I had to reinstall kohya because updating fucked it (of fucking course). But now it won't install cudunn for some reason?
>>16274 Hot to me
>>16272 Surprised they're still not using safetensors for this shit
(1.75 MB 1024x1536 00065-76050218.png)

(1.84 MB 1024x1536 00067-1447134763.png)

>>16267 Thank you for the update! As >>16268 said, it seems like I have to crank up the weights a bit to get the model to recognize the prompt. Other than that, this is definitely an improvement over V1. I could do more testing later if you're planning a hypothetical V3 at some point. >>16266 That's a nice size, too. >>16272 >It's not TTS so yes I did have to use my voice. Makes me wonder how many anons here have img2img'd themselves for their gens...
>>16279 there was the slutposting thread in /trash/ that spun off of /aids/ not the gayest thing i've ever seen on 4chan, but pretty close
>>16279 not sure what model or specific MBW preset I used in my mix that would make the need for emphasized prompts, share the catbox and I'll work on V3 not too long from now. I'm satisfied with V2 from now but once I'm bored enough and made enough LORAs again I'll work on fixing that for V3
(2.39 MB 286x258 lel.gif)

>>16272 What's the best way to keep track of voice gen tech, and where do you find models and/or tips for getting high quality output? I was watching the generals in /g/ for a while but they seem to have gone away. I have so-vits-svc-fork and a few of the models that were linked on the github like the Trump and Obama ones and so forth, but they all sound kind of flanger-y and not as good as the stuff I hear online.
(1.68 MB 1024x1536 00043-1661963118.png)

(1.63 MB 1024x1536 00629-1661963118.png)

(1.78 MB 1024x1536 00015-1661963118.png)

>>16280 If anything, I'm surprised I'm not seeing more RPing with genned images. I guess people who would be into that are already dabbling with AI chatbots anyways. >>16281 Thank you, and here are some catboxes for one seed: https://litter.catbox.moe/vjuylr.png https://litter.catbox.moe/rrok7c.png https://litter.catbox.moe/updgnx.png First image was genned on V1. Second image was on V2 with no changes to the prompt, and you can see how it ignores certain parts of the prompt. Third image is with weights dialed up to match details and composition from V1.
(1.60 MB 1024x1536 00143-2047220770.png)

>>16284 One benefit to LoRAs being seemingly overpowering on Based66-V2 is the naizuri LoRA works pretty well with it.
(3.15 MB 1280x1920 catbox_6mnavz.png)

(2.58 MB 1280x1920 catbox_4gozes.png)

>>16285 I decided to make some naizuri myself, but this is using based66v1 (because v2 was still downloading) I don't think I can go back to normal upscaling with how detailed the tiled controlnet outputs.
>>16284 >>16285 Thanks I'll work on Version 3 in the near future, it's good to know it's really good with LORAs I'll probably keep version 2 around myself to use for LORA example outputs.Back to LORA baking
(1.72 MB 864x1920 catbox_c4aehl.png)


>>16283 There's a few discord groups but they aren't worth joining unless you're really lost, though it did help point me in the right direction for RVC. For the most part RVC and sovits models are posted on huggingface but if there's a model you hear on youtube you can't find anywhere else there's a good chance they either asked the original creator or they made the model themselves.
>>16289 ahh, the rare double pussy
Finally an interesting post from h schizodiffusion general >So for the tests on torch2.1.0+cu121, it seems to be about 4-5% faster than torch+cu118. With batch size 2, DPM++SDE, 640x640 and hi-res fix to x2 (so 1280x1280) Torch2.1.0+cu118 firstpass and hi-res fix pass: 5.8it/s - 1.04it/s Torch2.1.0+cu121 firstpass and hi-res fix pass: (6-6.1it/s) - (1.09-1.1it/s) This on a RTX 4090, and also built my own xformers (following this https://github.com/AUTOMATIC1111/stable-diffusion-webui/discussions/2103) with that torch version The torch version I used (and how I installed it's) > pip install torch2.1.0.dev20230508+cu121 torchvision0.16.0.dev20230508+cu121 --index-url https://download.pytorch.org/whl/nightly/cu121 Also, for those who didn't knew (thanks anon on the last thread), xformers (0.0.20 if you build it yourself) is now determistic. So that's an advantage vs sdp no mem attention. Finally, added torch2.1.0+cu121 with my own xformers to kohya scripts (lora training) and it works as well, similar 4-5% bump in speed vs cu118.
>>16291 >xformers (0.0.20 if you build it yourself) is now determistic no way
>>16273 Yay! Thank you!
I tried Draw Things on a 2021 M1 iPad Pro, first thing that manages to heat it up a bit. Takes precisely a minute for 512x768, 25 steps of DPM++ 2M Karras on SD 1.5 (pruned and FP16) It's pretty damn cool, you can even use hires, TIs, LoRAs and Controlnet on it. Eats up the battery like nothing else but still.
anyone know where to get the baka loli model?
>>16292 nta, but yeah, it's actually as of 0.0.19 https://github.com/facebookresearch/xformers/releases/tag/v0.0.19 >fMHA/cutlass: The kernel are now deterministic
anonM, the preview image for based66 NEEDS to be tenshi eating a corndog next to the route 66 sign. I am BEGGING you
>>16297 this guy gets it
>>16297 Just an angel character or something? Unless there’s a tenshi LORA share it and I’ll work on it soon
>>16299 think he means kanata
>>16300 Ah that’ll be easy then
(168.12 KB 1024x1024 download (2).jpg)

>>16300 >>16299 you faggots cant possibly be this new, like are you kids legit zoomers or?
>>16302 >there's only one tenshi in the entire world kill yourself
>>16302 Ah that, I’ll probably do it for the preview image of V3 since that’s the final version I make before just moving onto 67
>>16303 >NoOOoOOo YoU CanT ExPEct MEeeE tO ReCogniZEe OnE oF tHe OlDEsT AnD MoST IcoNiC ImAgEE mAcrOS iN CHaN CuLTuRE NOoOoOo!!!!1!! faggot ass tiktok nigger vtumor zoomer lurk more or gtfo
>thread became schizo diffusion general now can never escape that fate even on dead ass boards
>>16305 >sage faggot >nooooooooo muh oldfag culture sorry, i'm not a pension age boomer and knowing every single 2hu image macro doesn't put food on the table now go back to studying COOKIE or whatever oldfag schizo shit you're into
>>16306 i mean thats kinda expected, if you dont gate keep and tell under age shitters to get out they will just start flocking to places like this like flies cause "woah look at me in part of the sekrit underground kool kool kids klub now woaah", 4chan devolved into reddit without login requirements ages ago and so will this place if people dont take care of it
>>16308 you do realize the only person who's acting like a faggot here is you, right?
>>16309 >immediately spergs and lashes out telling people to kill themselves when called out as a child >somehow views himself as the mature voice of reason in this situation
>>16310 >suddenly isn't ok with bantz
I honestly think the x2 4090 LORA maker that made 200 LORAs ia a Larper, they only post outputs and none of their LORAs at all, it just seems too fake to be honest
I guess based66 struggles when you started to put 3 loras. It does forgot the character outfit and didn't nail the gacha prompt concept. Granted, the loras are typical mill on the run civitai.
>>16313 If it's based off Civitai LORAs I can't really find it a reliable source of testing 3 LORAs, but I assume you were testing a concept/character/style LORA?
>>16314 Yeah.
>>16315 I can test out the Naizuri LORA with a style LORA i can quickly bake, main focus with V3 would be to see what model or specific block weight within V2's recipe is causing the need for higher prompt strengths but I'll make sure to test out 3 LORAs before I publish it.
>>16315 max strength of each LORA was at least in the 0.8-1 range correct?
(529.66 KB 512x768 00033-1661963133.png)

(531.68 KB 512x768 00034-1661963133.png)

I found the problem, it was the final model I added to the 6th part that was causing issues, I tested out Part 5 with a 3 LORAs that the final model failed to achieve but part 5 was able to achieve right away, the final model had no recipe so perhaps there was a finetuned model I was unaware of causing this issue. I'll see what I can do MBW wise to fix it to include the final model still or just leave it out completely to save the hassle, I'll test out if any MBW settings can help
I was planning to just go back to P1 for Based66 V2 and start from there again switching some models up but I realized if I did that I would just be changing too much and just making Based67 at that rate. I rather wait sometime before I decide to do that so I just worked on specific weights for Part 5 and Part 6 of the model to fix the 3 model LORA issue however I'll be sharing a pixeldrain link of the pruned version of V3 first so the anons here with more reliable testing can find any issues I'm not aware of (because honestly I don't collect too many style LORAs or do complicated outputs that require me to use high prompt weights) https://pixeldrain.com/u/hqDvXLpx
>>16295 Baka loli model? I don't rmb making one but my models are capable of lolis. only a few has my models as they're incomplete so I don't wanna release them just yet.
(3.82 MB 2196x3072 catbox_07qodc.png)

(3.98 MB 1464x2048 catbox_9aj5mn.png)

(2.91 MB 1280x1792 catbox_sqzit5.png)

(3.16 MB 1464x2048 catbox_i5ui4j.png)

Though the model is coming along quite nicely. let me know if its too realistic for this board it leans near 3d but I don't think its realistic.
>>16321 I wouldn't say it looks realistic but more of a 3d render feel to it. Still looks cute
>>16319 You probably can put this as V3 at least. I tried to mix 4 loras without latent or region prompt for giggles and it still works to some extent.
>>16321 no that's really nice actually, i really like the "soft 3d" look
>>16326 Yeah must have been the last parts causing the most issues, good thing I kept them o it was easy to fix (despite the large filesize)
>4060ti 16gb might be real It's over...
>>16330 did jewidia finally stop trying to psyop people (including themselves) into believing that faster vram is better than moar vram?
>>16331 >into believing that faster vram is better than moar vram? to clarify, muh faster vram hasn't been a good counterpoint on anything higher end than xx30/xx50 cards and even then i can't even recall which "modern" card last had regular DDR ram as the "lowest of the low scumbag penny pincer" option, probably the 1030 which is just the Nth rebranding of some 10 years old bottom of the barrel card
>>16333 there was a 1030 with gddr5 and a 1030 with ddr4. 65% less memory bandwidth and zero effort made to tell ignorant people what the difference meant. just so everyone remembers what a bunch of jews nvidia are.
>>16335 wasn't it ddr3? or maybe i'm thinking about the previous rebranding but yeah, exactly that. faster memory is completely irrelevant on anything but the lowest of the low-end like the 1030 and i'm not aware of any NEW offerings that use regular DDR ram so hopefully jewidia will stop grifting people with the "M-MUH FASTER MEMORY IS BETTER THAN MOAR MEMORY" but who am i kidding, i expect the 5070 ti to still have 8gb lol >>16336 >r*ddit dismissed on principle
>Inkcels got so powerful and had influence over the credit card jew companies that Fanbox is banning AI art
>>16338 once again, i guarantee you it's chinese spamming the service with actual CP and a handful of realistic gens dusted in there as cover
>>16339 you mean (((chinese))) right
>>16340 the cheating businessmen, the nasty landlords, the ethnic diaspora that seeks power and then only promotes their own, the fifth column that infiltrates every nearby society, the ones that operate their own illegal police forces on other people's land. the chinese.
>>16342 ah yes, the xijinbergs
>>16336 >use 30-100 images >style trigger words ain't reading any further
well at least the /g/ thread based on the Fanbox/Fanita AI Art ban shows that a lot of people still think AI art is retarded, AI artists are an oppressed race unironically
>>16345 I can't even stomach going to sdg now, it's absolutely unbearable
>>16346 I was mostly looking at the seethe posts regarding the 16gb 4060 but the AI image ban that fanita/fanbox rolled out is the most active thread kek
>>16344 I usually recommend 50-150 images, but 100% style trigger words are horrid
>>16328 indeed. soft 3d fluff
ooh sorry abt the name. i forgor to have it off.
>>16291 torch 2.1.0 doesn't work for me unfortunately, webui gets stuck at installing requirements (even after deleting the venv and doing it from scratch) So I'll just try building 0.0.20 on 2.0.0
>>16351 reporting back just to say it works
>>16352 Wait, it doesn't I'm getting different gens if I change batch size now
>>16321 while i like this style of images and do a lot of them myself i feel like you would need to be straight up delusional to not say that this isnt just hard leaning on the 3d/realism side of things wouldnt surprise me if this sort of stuff just gets nuked cause of the realism rules or if weabs start crying over this not being anime enough for them
>>16354 >i feel like you would need to be straight up delusional to not say that this isnt just hard leaning on the 3d/realism side of things nice tactic kike >>16354 >or if weabs start crying over this not being anime enough for them unfounded and irrelevant concern
seraziel anon are you alive?
>>16357 yes currently genning and posting fat guras on /vtai/
>>16358 need chubby shark who ate too much bathtub pizza also hi
>finally getting around to making my own LORAs >suddenly have to learn all this crazy incoherent machine learning jargon shit fuck man, I just wanted to make waifus
(153.77 KB 605x913 107141976_p0.jpg)

Someone knows what kind of model or Lora I could use to get this kind of slim, thin thighs, body type?
>>16355 go ahead then retard, start posting your 3d cunnies see how long it lasts
>>16362 dumb twatter-minded faggot thinks soft 3d = realistic = jailtime (also implying you'd need to care about what faggots think of your gens) what a faggot, i bet you think that you might as well use the no-no model at this point and that artists like sakimichan could get jailed for drawing realistic portrayals of canonically underage characters do me a favor and gargle on some industrial grade bleach lol
(1.56 MB 1024x1536 catbox_ixym3f.png)

>>16361 hmm, I don't know the model they are using, but I am almost 100% sure there is some mikozinn in his gens based on the eyes, I have reason to believe it may be using one of the loli diffusion models + a mikozinn lora at 0.4 or so, as I was able to get somewhat close using the (amittedly newer) loli diff 0.9 based65 model + a mikozinn locon I had lying around, I can't say for certain how to get that exact effect though, if it's not one of the lolidiff models, then it might be from another set of models, that's the hard part, I don't know exactly what they used. hmm, looking again at some of their more recent works... perhaps spring2013 is also in the mix. based on the skin texture... actually... the base model could be loli_A, which unfortunately is hard to get now because the huggingface for it got taken down and to my knowledge nobody has uploaded it again. I might do more digging because I like the style but I really shouldn't be going on huge tangents when I have the UI to make still lel
>>16364 cute seeing that with the olympics being 1 year away you are already so hard at work training for mental gymnastics
>>16366 pot calling the kettle a nigger you can't save face, kike, i know what games you're playing bring your concern trolling somewhere else
(761.07 KB 768x960 catbox_jwhwti.png)

>>16367 you are sounding like a clown anon
>>16365 >loli diffusion models Wait, those exist? Where do they get uploaded
>>16368 you're the entire circus, yid
(654.76 KB 796x1200 106137509_p0_master1200.jpg)

>>16361 Found the guy https://www.pixiv.net/en/users/90418429 I'm really curious about how he achieves that kind of petite body frame
>>16369 his most recent set was uploaded to pixai.art https://pixai.art/@josefjilek/artworks/models older version are also on his huggingface and proton https://huggingface.co/JosefJilek/loliDiffusion https://drive.proton.me/urls/6ECAX4KNDM#RFUtogGtSNtX also side note, this might actually be part of the base used for those gens https://huggingface.co/sleepotimer/Model_A >>16371 yeah, 100% the eyes give away that mikozinn was used in the combination for the model.
some really annoying faggots in the thread today
>>16371 >that thigh gap New fetish adquired, I wonder how good would it look on nonloli models
you all piss me off, all ai thread anons have started to do that, I guess artcels were right about one thing. That we all hate each other
>>16373 are you looking into the mirror over there bud?
>>16329 Good work man, I hope you can upload v3 on huggingface within a week. Otherwise, I probably need to repost it to my own for repository and easier fetch.
>>16378 oh you're a colab user? I thought they banned that shit already. Yeah I'll upload it I'm just trying to make the tenshi eating a corndog preview image for V3 kek
>>16372 The model A thing looks like a focus on 7th anime to be honest. The loli diffusion model is just a mix made from that mix using pre-existing models, I honestly thought it was a finetune
>>16379 There are some that get away with by changing the folder's name. Hopefully the banned issue will deter the phone collabfag at least.
>>16380 oh yeah, the loli models are mixes, but they are mixes with loras and other models that are specifically designed for making genning lolis easier. I know that they made a dataset for training a loli model, but I don't think they have the gpu time to spend on training a full model on it. Depending on the time I have, I might make a loli body type lora, because that is probably going to be a worthwhile thing to train
(99.12 KB 707x482 1683757862851903.png)

Wow thanks /g/ now I find out Patreon is nuking AI shit
>>16383 oof, first fanbox, now patreon https://official-en.fanbox.cc/posts/5934381 at least fanbox is a "for now" situation. we will see if they let AI artists make fanboxes going fowards
>>16382 is there a centralized place for loli models yet?
>>16386 I know loli diff guy created a website that isn't really a direct host to lora and models but they need somebody who is in a country that specifically allows loli to host it. beyond that, I just keep up with releases that people bring attention to in various servers and websites
>>16381 ok test it out, I'm just trying to make the preview image with tenshi before I upload it to Civitai, but so far the 3 LORA issue is fixed and someone tested out 4. I'm not sure if the extra weight strength is needed like V2 but if you can test it out that will be cool. https://huggingface.co/AnonymousM/Based-mixes/blob/main/B66%20V3%20P6-Pruned.safetensors
what settings do you use for hires? i use the same steps as main (25) and remacri at 0.4/0.3 or anime6b > latent @0.6 if i'm feeling lucky not really happy with the results but after trying out remacri, ultrasharp, animesharp, lollypop and all 3 ultramixes (restore being the best imho) i think remacri wins most of the time and upscaling is slow enough that i don't feel like cycling through all upscalers
>>16389 Remacri does not need more than 15 steps, so that will speed it up a bit. I also started doing my hires in img2img so I can try multiple seeds at a higher denoising strength. Here's a link to where someone goes into a little more dept: https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/2213#issuecomment-1274137775 What you need to know is that you can choose the default img2img upscaler in your settings.
>>16390 not sure if the resolution stuff is something i'm gonna bother with but i'll test the cfg scale trick and see if stuff comes out sharper
>>16391 I don't really do that resolution calculation stuff. You just have to manually set your target resolution in img2img. I think the CFG thing that dude is talking about will only work when you already have a low res image and then upscale it in img2img
>>16391 Either way, good luck! I hope some other people share their methods as well. We can always improve
>>16392 >I think the CFG thing that dude is talking about will only work when you already have a low res image and then upscale it in img2img i'm tired so i can't tell if it's placebo but after messing with it i think switching from cfg 7 to 9 does sharpen stuff up a bit though the composition also changes ever so slightly 10 and above start to look fried and 15 is way too much
>>16390 >Remacri does not need more than 15 steps is it something you figured out or is there some readme i missed somewhere?
>>16336 Actually looking at his settings, it's like a perfect recipe on how to deep fry your lora.
the sciamano240 lora from gayshit works well enough but it's pretty fucking hard to upscale, idk why but it just doesn't upscale well skill issue probably but i'm too tired to keep experimenting, oyasumi
>>16395 I think I read about it in one of the earlier threads when I asked about upscalers. But testing it out for yourself and your own settings is always recommended. You could easily make an x/y/z or two... right?
>>16388 Thanks, seems better now, kinda need to put a little strength on the prompt but that's probably the lora itself since the example images using that too.
>>16398 >You could easily make an x/y/z or two... right? i'll test it again tomorrow after i've rested my eyes but i don't use xyz plots, especially for small things like this i'm a "open in a photo viewer and alternate between the current/next pic as if i'm playing a shitty stepmania map" chad, just works better for me
>>16399 is it like 1.3 at least? One tester with V2 had to do 1.5 to get something right
There was a "sex education" lora that got taken down from civitAI before i could download it. Does anyone have it or know where else it might be uploaded? It was for putting an embarrassed girl with her ass out in a class room with the teacher pointing to her.
>>16401 1.2, no biggies.
>>16403 Ah that's good then, with the I assume your Sweep Tosho LORA was by the Civitai user with the Special Week anime avatar? I was testing their satono diamond LORA for a 3 LORA test (naizuri and a style LORA I made) and it works fine compared to V2 so that means that I fixed the issues. I just have to make the tenshi corndog preview pic but even with Controlnet AI still hates finer text kek
Does Deviantart at least allow loli, just in case even art sites like Pixiv ban AI image uploads?
>>16400 >i don't use xyz plots, especially for small things like this Wait what? But Anon, xyz plots are particularly useful for easily generating multiple images only changing a select few variables!
>>16389 I'm lazyfuck that rarely use regional prompt and keep using fried loras. But SwinIR 4x probably the most tame for me, I can throw everything easily and it usually just werks.
>>16371 that's a nice style. i feel like that's a specific artist that i can't put my finger on, a blue archive guy mostly i think. the face on this one. plus some yassy.
>>16387 are there a lot of countries where loli hentai is illegal or something? it's legal in the US so that seems like a pretty obvious one....
>>16402 civit is seriously starting to get on my nerves with that shit
>>16411 it's a legal grey area in the US, as in it's legal, because there's nothing saying it's illegal not because it is explicitly said to be legal, which is why the US isn't the best choice, I'd think at least. I don't actually think there are a lot of countries where it is illegal, but likely they are being cautious. >>16410 I am also pretty interested in finding the style for this, because I certainly enjoy it as well. seems like this may become a board effort to replicate lel
>>16410 body type reminds me of the artist hagi a bit, but I could be just going schizo
anyways, closest I got using the loli model posted earlier. I used a mix of the mikozinn inspired on civit, wagashi v3 from some past thread since I thought the ass reminded me of wagashi, and ootomo to correct the eyes a bit. First two seem to do the heavy lifting, but wagashi throws off the eyes so I'm not sure. Skin and definitely hair texture could use work. Things look nowhere close with front shots. Also, her dumb smile. The thin limbs, I unfortunately have no clue. I'll try some stuff tomorrow. Counter intuitively think it might be one of those semi real models in the mix.
>>16360 You don't really have to. Just take a lora that you like and look a the settings it was made with. Nobody here actually understands what is happening.
>>16390 Where did you get 40hara lora?
>>16354 Hi. If you look at the metadata of the models, some of them have the word "Outxattn" in them. I have used https://github.com/rohitgandikota/erasing to remove the concept of "realistic" from the model. By the time it is released, when you type out the word "realistic", the model will ignore the token. I have made the text encoder unresponsive to the token as well as weakened the realism from the model's OUT layers. I would go to great lengths to censor "realistic" from my anime models, but soft 3D should be fine. Baka-Diffusion will not tolerate any realism in its newer models. But if you start whining about soft 3D, then I can't help you.
So... when tagging datasets for an art style typeLora, which tags do you remove if any. Also should I had blacklisted the "mosaic censorship" tag?
>>16413 yeah its legal as long as it passes obscenity test its kinda a grey but people have been put in jail for it. https://www.mtsu.edu/first-amendment/article/1585/miller-test https://en.wikipedia.org/wiki/United_States_v._Handley
fml >>16420 meant for >>16411
>>16419 No, for censorship. You either put mosaic and censorship on negative or time to uncensor the dataset by yourself.
>>16294 iPad anon, share your thoughts on 512x768 image size. I use the same iPad and whatever model I use it looks horrible unless I manually go for 640x960. Wonder if I do something wrong.
>>16415 catbox?
>>16407 yeah but it's easier for me to just alternate between two images at the speed of light, i can never tell small details apart on xyz graphs
>>16423 Mine is the smaller 11" Pro, you might have the 12". That one's higher res too iirc.
Kohya had some pretty cool updates with his scripts finally, not sure if Easy anon wants to deal with some of the experimental stuff but Kohya added the option for the preview images to have hires fix with the latest update
>>16427 I'll probably mess with it once I get the base UI done. as it turns out, fitting about 90 arguments into a one page UI that looks nice is pretty god damn hard.
>>16428 tabs?
>>16429 I've decided to use dropdowns so that the user doesn't have to go through tabs. I'll probably share what it will look like once I am done enough to put it all together
which one of you fuckers broke containment on the based66 page
>>16430 tabs are probably fine, dropdowns would be a nightmare if you're gonna implement tooltips it doesn't have to become aurora 4x levels of UI
Based anon disable comments next time
>>16433 Usually comments don't end up with garbage like this, must be some schizo diffusion resident that knows about this place
>>16434 Can't you disable them now
>>16435 Yeah doing it now, guess that's what happens when we have more schizo residents here now
>>16436 Anyways comments don't matter much on Civitai anymore, those normie AI users would probably beg for a VAE baked version of the model or something
>>16433 >>16434 based66 is supposed to be more 2Dish than b64 right? I didn't like b64 because of its 3d render feel and noisiness
>>16438 Yes basically, it was supposed to be based65 but with LORA compatibility fixed and sharper lines. My only issue with 64 was honestly how kl-f8 made it look more orange but with version 3 (hopefully the final version) it does well with multiple LORAs and no weird issues with stronger weights with prompts when usings LORAs (which I found out the reason was due to the final model in the mixing recipe had a finetune within it)
>>16439 >kl-f8 Damn I didn't even think of trying it out with different VAE. I noticed the color tint as well. Thanks I will try b66 out
>>16432 I do plan on having tooltips, but I'm using labels and stuff to explain what each input is without having to mouse over it of course, I just didn't think tabs would look nice. Main reason why I want to use dropdowns is because I'm separating out the args into groups and I can use the dropdowns to nicely separate them out. that being said, final design is pretty much not decided yet, I might pivot into using tabs if it makes sense, I'm sure you will see me asking for feedback once I have enough done
>>16424 https://files.catbox.moe/ffqieh.png https://files.catbox.moe/t1x3wf.png if anyone can figure out what artist(s) or models the sharp eyes with eyelashes come from, the thick line weight on the hair, and the thin limbs, it'd be progress. I've noticed some particular patterns in the armpit's i've seen in my own gens so I'll be looking into that a bit
Sword deepthroating lora that got shoah'd https://files.catbox.moe/e483ju.safetensors no specific trigger words, but "oral, hetero, fellatio, blowjob, deepthroat, testicles, penis, huge penis, standing male, squatting female, arms behind back" work
>>16264 i don't look at this place for a week and now we have retarded namefags like "BakaMan" and /sdg/ tier clown posters. what happened?
>>16349 did you ever upload it anywhere btw?
>>16444 More people found out about this site, sad it didn’t lead to more active posters however
>>16418 would like a link for the model from before this faggot >>16354 complained about it, i love the style exactly as it is here >>16321
>>16447 please wait for me to refine it alittle more. and I will give it to you.
>>16448 fuck off attention whoring namefag. this is turning into a repeat of 4ch /hdg/
>>16448 appreciate it, thanks
>>16449 name gets saved with cookies. thanks for the reminder it shall be deleted. also stop whining please its not like im cumming over your face.
This thread started to become shit once people like Based anon started to avatarfag, literally it’s the reason why people find this place but they’re fucking schizos
>>16452 are the avatarfags in the room with us right now anon?
>>16451 >make one post to tell you to fuck off with your bullshit >please stop whining!!! lmao please, glad you weren't insufferable enough to keep at it but feigning ignorance is laughable >>16452 based anon is the only one that makes models i'm really interested in and rarely talks about basedmix except when a new mix is coming, so i'm fine him
>>16454 Still a problem when we have people finding out about this place because of faggots like him
>SD understands "grey blue eyes" >it doesn't understand beige shorts
>>16452 Well at least Based Anon is doing stuff for the community with merges unlike your nigger ass is.
>>16454 not advertising my model here. someone asked so I replied. now my whining sentence is valid so again anon. please stop whining its not like I cummed in your ass
>>16456 ah i see it's "sd ignores half my tokens" time again
>>16457 All his mixes are garbage Based66 is as shit as Based65, bet 4chan would agree
>>16460 why are you even here then. seems like you hate literally everything about this board. why not go back to 4chan where everybody likes to complain.
>>16460 haven't tested 66 but from what i've seen it's not as good either seems like b64v3 was lightning in a bottle
>>16461 Niggers made this place easy to find so blame it on the random advertising posted on 4chan >>16462 Literally, faggot thought he could make another good model but he’s getting btfo’d by Civitai merges of all things kek
>>16463 I tried dicking around with absurdly massive hips and short torso in controlnet but unfortunately it wasn't that convenient.
she's even flipping me off... what the fuck...
>>16466 LMFAOO
>>16445 SauceNAO should find my pixiv usually there and the same name on twitter.
>>16466 yeah idk, today's not a good day i guess
>>16463 the big toothy bean smile is probably a hint, not a lot of loli artists do that without a fang or something
>>16420 not concerned about it, but I see the need for hosting somewhere a bit more liberal. Still feel bad for that poor dude having to deal with those post office/court bumpkins
>>16427 i gave up on kohya due to poor support for <=8gb vram cards that should have no problem training loras if he would just implement the right settings or make them available in the UI. moved over to use the a1111 extension instead for my lora training now
>>16472 yeah youll be fine as long as is its not something like "loli gets fucked in every hole by fat ugly bastards in a public bathroom then gets pissed on" or shit like that,
>>16473 wait you can train LORA on the webui now?
>>16473 Doesn't the extension just use the same backend as the extension? As in, sd-scripts?
>>16463 What's the exact promts/model/Lora on the first picture? I has the perfect body proportions for a character I wanna try working on
(763.97 KB 768x768 00149-1689216865.png)

There are a bunch of Megumin LORAs on Civitai and I got no idea which is the best one. The one in the gayshit repo is kinda wonky
>>16477 https://files.catbox.moe/edmv0i.png yeah sorry for the mess in my catbox prompts. Too lazy to move all the loras to addnet folder and I honestly don't have too much faith in prompts making the biggest difference past a certain point so I tend to leave in stuff from what I was working on before. I used a simple controlnet to drag her arms up top.
Wow Asus fucked hard with their motherboard voltage issues and voiding warranties with their "fixed" beta BIOs, /g/ is seething and youtube is seething too damn
>>16480 Guess who's asking? Hi, it's me. I still haven't built that 7950X rig two months later.
>>16481 The bios that was supposed to fix the SOC voltage issues blowing up X3D chips and that one 7900 is a beta and using it along with EXPO will void warranty so Asus is currently telling people that install the bios that is supposed to fix the EXPO over-voltage issues will void your warranty if you use it.
>>16482 I know. Like I've said, I'm the guy asking on /g/.
>>16454 that's 2 whiny posts now
>>16460 so annoying
>>16464 this aint your clubhouse. go host your own site faggot
>>16479 How strong is your prompt that it eats all of my character specific Loras lol
>>16487 lmfao, I probably use way more parentheses than I should. I do also find myself having to be very specific with character traits (hair, eyes, breasts) with character loras
(824.37 KB 768x960 catbox_6qi9e4.png)

>>16485 >>16486 you sound angry, you need some balloons to make you smile so much clowning in these threads lately, what happened
>tried a nichijou lora on a realistic model out of curiosity
>>16490 >"here's what happened"
>>16491 i don't want that evil on my computer, try it yourself if you're curious
>>16490 >realistic model >>16492 >evil could've just said you tried to use it on the forbidden model
>>16490 instant kigurumi
>>16493 >hurr durr here's some gay assumptions gonna post some realistic nichijou clowns if you keep this up
(3.12 MB 1280x1920 catbox_5qmtbv.png)

(2.75 MB 1920x1280 catbox_fr3noy.png)

(3.29 MB 1920x1280 catbox_sm4h4m.png)

I wish my tail hug lora worked better
>>16496 >human gura been a hot minute since I've seen that, considering I've been prompting her as a sharkgirl with cock and balls my dick is too destroyed to feel anything right now though
>>16497 seraziel anon is that you
>>16495 its ok anon we get it we get it, you got your cunny stabed big deal woah you are so cool
shit for some weird reason /its is running slower now at past resolutions that didn't give me trouble. Oh well, updates do fuck some things up I guess (LORA training)
Pixiv follow the suit, eh. I thought it was only fanbox and patreon.
>>16501 >Artist charges over 30,000 >AI artist can do the same work for just 3000 Kind of makes sense ngl, it only went like this because we were such cheapsakes when taking requests
>>16501 It's all going down so quickly, maybe they were just waiting for one company to pull the trigger before everyone started doing it
>>16503 They can't stop the prices on art plummeting. People will just paypal AI artists if it means saving enough money.
>>16498 yes furry futa is just so much easier than regular futa with the furry models so I can finally coom to very good futa with that lora next up ass shots
>>16504 You presume that paypal will allow such transactions. You already lose your account if you're discovered to be selling NSFW materials. Couple that with the increasingly tenuous legality of AI generated images, you're going to have a problem on two fronts.
>>16506 I seriously can't believe they will be able to suppress market mechanisms with their astroturf but we'll see.
>>16502 Well, in the end money triumph over such a trifle such as artistic integrity.
>>16463 Can you point me in the direction of that Mikozinn lycon?
>>16493 >forbidden model go back
>>16509 NTA but assumedly he's using this one https://civitai.com/models/17291?modelVersionId=20431 also side note, lycon is actually locon, but that's me being pedantic, so you can ignore that
>>16507 Please use less buzzwords, it makes you very hard to take seriously and makes me heavily question my pariticipation in this.
>>16513 what makes you think anyone wants you participating here newfag
(2.04 MB 1464x2048 catbox_henwaf.png)

(7.10 MB 3294x4608 catbox_uiu2hk.png)

(2.93 MB 2196x3072 catbox_x1zq3p.png)

(1.66 MB 1464x2048 catbox_i7a6xy.png)

WIP Fkey LoCon
>>16515 looking good man. i love that artist.
>>16505 send some
>>16513 "astroturf = buzzword" go back newfag
>>16506 i wouldn't sweat it at all honestly. there's so many ways to skirt blocks like paypals (i bought a sexdoll using my paypal account. the vendor just tells you not to mention anything sex related when messaging them). this is a young person tech for the most part, and while a lot of normies are dumb af, they'll be savvy enough to use coinbase/crypto if it means saving money, especially in this economy and bleak environment. old dinosaurs with too much money can kick and scream all they want trying to protect their revenue streams, but at the end of the day there's going to be an absolute glut of AI generated stuff which is going to crash the market regardless of whether it's done by new businesses or whether it's done "underground" on reddit/chans/elsewhere. you'll also have a cat and mouse of people creating AI art that evades detection. convenience is king in a consumerist economy, and this is just way more convenient in most cases. i can generate keepable hentai images faster than i can download ones that are worth keeping with 5 year old hardware
>>16510 there's a forbidden model? do tell
>>16513 Fuck off
>>16517 done
>>16522 ... where are you posting them again?
>>16524 based
(1.70 MB 1024x1536 00027-1682865255.png)

(1.52 MB 1024x1536 00137-2896895453.png)

(1.67 MB 1024x1536 00113-4251897131.png)

(1.82 MB 1024x1536 00029-2051921085.png)

>>16520 https://mega.nz/folder/ctl1FYoK#BYlpywutnH4psbEcxsbLLg added maydaykiki. based66 is out? neat. i'll have to give it a try
Why does upscaling take up so much GPU power holy fuck
>>16399 Catbox me on that. I've been digging zoom layer / multiple angle but I want a consistent upskirt one.
>>16526 Don’t use it, it’s bad, so bad to the point the anon that made it had to lock down the discussion for it on Civitai kek. Based64 was a lightning in a bottle situation that should have stopped there
>>16529 mhm tell me more.
>>16531 It’s just not that good, had troubles with prompts when using LORAs and more than 2 LORAs with the version they released first, faggot made some V3 shit that said it fixes those issues but I could care less. Feel bad that HLL’s anons finetunes got used in such garbage mixes
>>16511 this anon got it
>>16529 >>16532 you're trying too hard
>>16534 Just hate anons that publicly expose themselves, it’s Reddit behaviour
>>16529 why wouldn't I use it? the characters (especially hips) in those look great, it's just the images are a bit too grayish
>>16536 i'm not the shitposter but that's not based66. that's just 64v3 with the maydaykiki lora listed
>>16526 got distracted. which one of those is the one you used for your images?
>>16537 nm, i tricked myself into thinking it wasn't maydaykiki when ctrl+f didn't show search results. that was weird
I noticed some metadata of Civitai LORAs showing that they are being trained at 640x640, I thought that type of training was only being posted on 4chan until they finally started testing out 1024x
>>16531 why are you acting like what that anon said isnt factual? you can go to civitai yourself and check it lmao kid is desperately trying to censor anyone that criticizes his nothingburger models, this place is more of a circlejerking echo chamber than reddit nowdays wtf happened
>>16543 Based facts, imagine closing down the discussion section if your model on Civitai, only Zhangs do that holy kek
Bros, I feel like I missed the boat somewhere on how to get decent hands especially when using loras. I feel like badhandsv4 and bad-hands-5 if anything just make things worse. Any tips on how to prompt better hands?
>>16545 it's less about your prompt than it is about how much your model clashes with itself and your loras and your samplers. just stick hands in your positive prompt and play with it for a while. it's also about euler a, 4x ultrasharp, <lora:zankuro:0.2> <lora:rustle:0.3> <lora:typo:0.3>
(89.67 KB 1178x883 86ui2fdt1b358nh_50.jpg)

Is there any detailed spread pussy Lora out there that has at least grade A anatomy knowledge about how a coochie looks? If I make one how should I tag it
>>16515 I made one a month earlier if you'd like to compare, it's probably shittier though https://mega.nz/folder/oUokGaJR#OWRSRLgTrr9MrbPzkAFfYA/folder/4Ih2iZDR
(1.11 MB 1024x1280 33345.png)

How to achieve this simple color style
(21.85 KB 936x360 capture2.PNG)

>>16425 Yeah, that's understood! But the xyz plot script makes it way easier to generate the comparison images without having to manually change the settings for every image you gen. You don't need to actually look at the xyz plot to compare
>>16417 Getting back to you late, cool that you can see it's 40hara! It's a lora that only really works for the maid as it is trained on 14 images only of Itou Chitose. I got it on 4chan from RandAnon
>>16548 Looks pretty good too man
What is this multres noise thing with a 0.3 value? I don't like things being added to LORAs that I don't understand just in case it ends up being like noise offset
>>16518 Tell me, when have you used "astroturf" outside of this one particular context. I've never seen the word thrown around with so much abandon until very recently, and only about this. You can get snippy with me all you want, but I'd like you and the rest of the diaper babies to practice some introspection and think about when you actually started using that word with such regularity and what might've inspired the first recent usage of it.
>>16547 this man asked a pressing question dammit, and by god he needs one!!!
>>16555 would you say use of the word is being astroturfed
>>16555 I'm going to ban you if you don't knock it off. This is your first and final warning.
>>16545 Legit, I think your gacha chances of getting good hands is higher if the LORA you used to train with had a lot of well drawn images of hands in them, if they didn't or they're out of view then it'll just default to the NAI hands. Don't quote me on this some LORA expert can correct me if im wrong
>>16555 there are literal policies/laws that are get this...DECADES old prohibiting obvious uses of astroturfing that are written using this term that you are claiming is some "buzzword" that people came up with last week or something its not our fault you are fucking 12 with a vocabulary consisting only of shit you heard on your apex/fortnite matches
Apparently there's more shit?posts going around saying having a LORA work at weight 1.0 is bad because it means the LORA is underfit
>>16555 >Tell me, when have you used "astroturf" outside of this one particular context Political contexts, advertising contexts, discussions about music production/audio equipment, car talk (god I hate mx5s) etc etc
>>16561 >having a LORA work at weight 1.0 is bad because it means the LORA is underfit great bait honestly
dropped out of LoRA training to hack together a better UI for ComfyUI. has a prompt queue now and I'm adding a in image editor with layer support
>>16564 >he made it look like gradio oh no no no no
>>16561 >1.0 is bad because it means the LORA is underfit nigga what
>>16565 I was just lazy with the theming, I'm making an alternate theme also. Mostly concerned with how it works right now
>>16566 It's just retards that think you can't go above 1.0 weight, so they believe it needs to be best under 1.0 so you can "see the ceiling"
>>16561 this is why anonymous boards are almost worthless for learning shit because half the time it's just trolling
>>16564 honestly, i would just skip adding an inbuilt image editor, it's going to make the UI way too heavy. it would be better instead to just have a button to launch with your photo editing program of choice and then to have the image update in the gradio UI when finished editing. no need to recreate the wheel with an image editor
>>16564 >prompt queue Nice, I've been wanting something like this together with a controlnet queue. Procrastinating on making it myself.
>>16570 forgot to add to my post: does anyone here have a lora for making flat chests? not seeing a usable one on the gitgud nor on civit
>>16570 I only opted for an image editor because it was an off-the-shelf library I didn't have to build from the ground up. It isn't always visible either, just when you open a dialog box. I was thinking about integrating krita just like you were saying though, I have nothing against it
>>16572 do you really need a lora for that prompt: (flat chest:1.3) negative prompt: (large breasts, huge breasts, milf, plump, curvy:1.3) wait the danbooru tag for milf is "mature female" what a bunch of fags i've been doing it wrong for months
>>16571 The way ComfyUI queues prompts means any arbitrary workflow can be queued, including utility things like previewing an image or applying a controller preprocessor, it's very convenient
>>16572 >>16574 you don't really need to increase the weight, just "flat chest" works fine
>>16569 sad part is that most of the time it isnt even trolling, its just retards saying retarded shit
>>16576 depends on model and loras. personally i enjoy using titcow artists to generate lolis and vice versa so sometimes you really gotta hammer it.
>>16576 >>16577 Is that simply just the fiz rot lora?
>>16580 fizrot+csrb+takorin @0.4 each it's my default style mix now so maybe that's why flat chest just works™
(881.73 KB 640x960 catbox_lx24li.png)

(742.78 KB 640x960 catbox_ajv5l8.png)

(4.19 MB 1500x2105 99635556_p1.png)

Anyone played around with the new controlnet preprocessor yet? No model needed just throw an image in and choose the reference only preprocessor. https://github.com/Mikubill/sd-webui-controlnet/discussions/1236
>>16583 Got a link for that csrb lora? 2000 fuckin style loras and I don't have that one, fuck's sake
(622.90 KB 640x960 catbox_yylraa.png)


(583.70 KB 640x960 catbox_3x73gk.png)

(790.73 KB 1346x1282 96148783_p0.png)

>>16584 On second thought it's neat but it's a better time investment to just make a lora especially since each gen takes at least twice as long. But for experimentation with a style it's ok
>>16586 do you need to update the webui for this or just the controlnet extension?
(1.87 MB 892x1345 111.png)

(1.78 MB 885x1343 222.png)

(1.48 MB 888x1348 333.png)

(1.51 MB 897x1338 444.png)

>>16371 0.3 fangdongye 0.3 miko on 3A last one is 0.3 fangdongye 0.5 miko 0.5 hagi didn't have the patience to gacha one like his prompt is just standing, ass, hip focus, (loli, child, petite, completely nude, spread ass, grabbing own ass, cum in pussy:1.3), pussy, nipples, small breasts, full-face blush, blush, anus, from behind, from below
>>16587 Just the controlnet extension
>>16584 I'm enjoying this preprocessor since now I can replicate styles of some AI pixiv users who don't share metadata w/o having to also bother training a LoRA for them
(1.18 MB 768x1152 catbox_1flcal.png)

(1.19 MB 768x1152 catbox_xfhtr0.png)

(1.19 MB 768x1152 catbox_i16d9r.png)

(5.10 MB 1536x2304 catbox_iwbdhc.png)

bunnies
(996.47 KB 650x983 22.png)

(1.55 MB 888x1325 2.png)

>>16371 same prompts/loras but on 1a ignore the broken anatomy it's the model over all i think
(1.35 MB 1024x1536 catbox_7aa3kd.png)

(1.76 MB 1024x1536 catbox_5v3a7b.png)

(1.26 MB 1024x1536 catbox_8o70ha.png)

(1.76 MB 1024x1536 catbox_72iv8s.png)

4chan said there's a cunny model released on Civitai, someone archive it quickly if it's a loli model
>>16588 do you have a link for the hagi lora?
(1.62 MB 1024x1536 00013-395118629.png)

(1.58 MB 1024x1536 00173-3123960119.png)

(1.70 MB 1024x1536 00012-1521550274.png)

(1.78 MB 1024x1536 00084-797165021.png)

(1.60 MB 1024x1536 00015-3906895280.png)

(1.68 MB 1024x1536 00048-1276469909.png)

(1.70 MB 1024x1536 00012-2120064452.png)

(1.78 MB 1024x1536 00084-797165021.png)

>>16598 use at 0.3 to 0.7
>>16573 i didn't mean integrate, i meant open externally. integrating it is just going to make webui run slower
>>16596 catbox for #3?
>>16594 We talking about KittenChow? That's been around for like a month and a half. It's not super prone to NSFW, so it's probably not going anywhere. If there's another one, I'd love to know about it...
>>16585 sorry anon i went to bed https://mega.nz/folder/33JBHQQZ#UvpkVuWwT43p8j37OsyWog/folder/WypwjDZA kinda funny that i fished it out first try from hundreds of mega bookmarks, didn't see it on there, went on gayshit and ended up on the same exact mega
>>16596 megumini is still bad and outdated :(
>pikamee is back >as a vwhorejo i hate being right time to gen pikamee until i forget about this
>>16549 Hey anon I replied to you on cuckchan didn't you see
>>16606 I knew it, disappointed as fuck I'm just going to take a longer break from vtubers to be honest
>>16601 https://files.catbox.moe/dr0chl.png the folder with the lora has all the preview images with the metadata
>>16603 it's ok, I'm an idiot and I did in fact already have it
>>16598 thanks dude
>>16604 one day anon
>>16545 gib Peni lora?
>>16549 Kagami lora?
>>16607 I did, thanks anon
(557.89 KB 870x1305 108051173_p1.jpg)

Style detectives how do recreate this style?
>>16614 Lucky star lora from civit.
There was a Liru LoRA that I had opened the tab for earlier and meant to download it, and then I went back to it was removed. It looks like the user got banned. Anyone happen to grab it? Not a big deal since I could always try training later myself. https://civitai.com/models/63745 https://civitai.com/api/download/models/70598 (also doesn't work now)
>>16584 The amount of people asking for diffusers support in that thread is funny to me. Author just doesn't give a shit and does what he wants
>>16618 i have a wolf girl with you lora, trained on the 3dcg game by that name made by seismic, which i downloaded in february. i have no fucking idea what the model you posted is supposed to be given that the only thing named "wolf girl with you" is the 3dcg game and the anime liru comes from is called something else. so this is probably not what you're looking for, but here it is. and he trained it on a furnigger model rather than anything base so it was probably bad anyway. https://files.catbox.moe/js92cf.safetensors
>>16620 Yeah I have this one, indeed not what I was looking for.
(1.45 MB 1024x1536 catbox_zijj4a.png)

(1.78 MB 1024x1536 catbox_krwfr3.png)

(1.89 MB 1024x1536 catbox_xcb6fv.png)

(2.06 MB 1024x1536 catbox_stkwsd.png)

(2.77 MB 1024x1536 catbox_vsk4pz.png)

(2.70 MB 1024x1536 catbox_5gsusw.png)

(2.50 MB 1024x1536 catbox_jnns40.png)

(2.23 MB 1024x1536 catbox_4wpa2c.png)

releasing this Ijigendd LoCon I baked while taking a break from the UI. it works pretty well, looks more accurate to one of the two styles at 1.3 strength, but looks like a combo of both styles at 1, which I like, and think looks really good. links: https://civitai.com/models/66180?modelVersionId=70833 https://mega.nz/folder/CEwWDADZ#qzQPU8zj7Bf_j3sp_UeiqQ/folder/7B5Qyaxb
(2.51 MB 1024x1536 catbox_krd8qm.png)

(2.25 MB 1024x1536 catbox_venejf.png)

(1.97 MB 1024x1536 catbox_ao2p8f.png)

(1.95 MB 1024x1536 catbox_h2jsv7.png)

>>16623 more previews
(848.90 KB 1024x1536 00026-2547202802.png)

(1.49 MB 1024x1536 00025-4105654901.png)

(1.65 MB 1024x1536 00080-4003475690.png)

(1.75 MB 1024x1536 00040-578619480.png)

(1.35 MB 1024x1536 00059-2802436990.png)

(1.47 MB 1024x1536 00032-3117908752.png)

(1.64 MB 1024x1536 00014-1169828239.png)

(1.55 MB 1024x1536 00258-215066906.png)

>>16622 Bear Man. pls gen some bear girls in flat color
what value are people using for the wd 1.4 tagger
>>16629 Cheers, anon. I will find out shortly if you bamboozled me or not.
(738.91 KB 640x800 catbox_zysvm4.png)

(863.20 KB 800x800 catbox_jkwcng.png)

(707.92 KB 640x800 catbox_zwy6f2.png)

Testing retraining an older unet-only lora. There's one pic of Nami in the entire dataset...
>gitpull because the new update lets you use the webui offline >ends up breaking my webui extensions and I can't output multiple generations anymore I always hate dealing with this shit
>>16613 It's just the shitty overbaked one from civitai, takes some gatcha and the dynamic thresholding helps.
(1.15 MB 1024x1536 catbox_v40czm.png)

>>16588 >>16592 good snooping dude. i think i added a bit.
>>16606 >lora of her already made I hate kikeson a lot but whatever I'm just hoping she can play whatever she wants there and none of the troons attack her. I feel bitter but I just hope she's happier https://files.catbox.moe/xsj2is.safetensors "henya the genius" is the activation token
>>16633 As the one who made that PR I can tell you it is unrelated to whatever problem you are now having
the new reference_only controlnet should not be combined with self attention guidance
>>16638 yeah I had mentioned this in the now locked discussion thread on github. I want to try looking into why exactly this is happening later.
tried out cu121 and it seems to not output anything aside from a CUDA error, must only work with the 40xx series
Can reference_only works from manga panel?
Has anything noteworthy happened over on 4chan in the past two months? Haven't had to time to keep up with how fast everything moves there
>>16642 besides catboxes and the occasional lora share, not particularly
>>16640 Can't even go past installing requirements on a 20xx
>>16606 Why redebut? I thought she quit
>>16645 trannies are probably too dumb to realize it's the same person
>>16646 trannies?
>>16646 She got bullied for HL so she redebuted to lose the trannies? Seems like a massive overreaction so I assume there's more to it. Real shame it's with that disgusting agency though.
>>16647 she tweeted about wanting to play hogwarts legacy trannies didn't take this well at all >>16648 idk, some can be very dedicated just look at how one tranny rapist (liz fong jones) can bend net neutrality to take proof of him being a sex offender offline, by harassing and threatening families of higher ups.
>>16649 >just look at how one tranny rapist (liz fong jones) can bend net neutrality to take proof of him being a sex offender offline, by harassing and threatening families of higher ups. chink tranny doesn't have that much power no matter what jewrsh wants you to believe he's just a mascot
>>16650 I hate trannies more than I hate josh and his relatively mentally challenged website
>>16651 i hate both equally
>>16645 >>16648 >>16649 Most reasonable thing that I can think of that Gyari and VOMs wasn't given enough support that she wanted and knowing Kson was her friend + knowing kson likes putting shit into JP indies heads she probably poached her into leaving VOMs and joining Vshojo by telling her they would do what she wanted, the harassment from trannies because of Hogwarts just made that process move forward quickly, not that Kekson was actually telling the truth given that tons of founding members from that company left because they weren't giving enough money and the fact Henya joined a company filled with trannyfans that harassed Silvervale for playing Hogwarts doesn't let her do "whatever she wants." I just find it stupid and disappointing but whatever, she's back and she's living by her lolicon tastes looking at her new avatar so it's ok at the end of the day.
Does anyone have the fisticuffs lora? The mega got taken down
>>16653 kekson is a fat spiritually jewish hapa who'd happily groom pikamee into opening an onlyfans, just give it time
>>16654 https://files.catbox.moe/fnrutq.safetensors here's it is, never downloaded a lora so fast in my life
>>16655 Kikeson is a fat shit, I already have a feeling she's going to push Mirai Akagi to join vshojo but whatever the only thing I can hope is that Henya's Loli avatar can be a deterrent for trainnies that harassed her in the past so she can play Hogwarts Legacy
>>16657 i'd still rather have no kettle than vwhorejo not-kettle tbh i wanted her back but ESPECIALLY not like this, someone wished on a monkey's paw
>>16643 Nice, thank you.
>vtuber fags stop simping to women just because they have an avatar. they're still 3d whores
>install new webui >old seeds look different uh is this xformers fault
(2.11 MB 1024x1536 catbox_a0akmn.png)

(2.33 MB 1024x1536 catbox_pkl69w.png)

(1.57 MB 1024x1536 catbox_ev2wzk.png)

(1.31 MB 1024x1536 catbox_h57gtg.png)

>>16661 I only gen them because they usually have usable designs and loras, and because /vtai/ is one of the least shitty AI threads, with various requests so you can crutch if you don't know what to proompt don't really watch any though
(2.33 MB 1280x1600 catbox_1wwfy2.png)

>>16664 Agree, only ever watched one or two streams but I only really like her voice and design. Doesn't make it much different from any other anime character to me
>>16663 xformers is now deterministic, so yeah probably
>>16659 How strange. The link told me it was taken down minutes ago.
>>16661 t. genetic dead end
>>16661 I've visited vtai and it seems like no one is actually watching that crap there. Some characters have attractive designs so it's fun to gen them.
I just hate inkcels getting all haughty just because a lot of sites and companies are starting to ban AI image tech
>>16670 i don't pay attention to them, personally
>>16670 >I just hate inkcels same
>>16672 was this an artist that went menhera because of AI?
(2.68 MB 1280x1792 catbox_kz382i.png)

(2.74 MB 1280x1792 catbox_utneni.png)

(1.25 MB 1120x1792 catbox_45ldcj.png)

(1.35 MB 1584x2048 catbox_h95adc.png)

Fkey LoCon release. (27mb) Can be found on CivitAI under Hosioka. Direct Link | https://files.catbox.moe/8w5yk0.safetensors
dead as fuck
(1.63 MB 1024x1536 00083-634843142.png)

(1.54 MB 1024x1536 00037-668213438.png)

(1.77 MB 1280x1920 00012-544624320.png)

(1.86 MB 1024x1536 00049-569136628.png)

(722.96 KB 1024x1536 00055-3410734345.png)

(1.01 MB 1024x1536 00267-1398225306.png)

(1.39 MB 1024x1536 00001-3081723956.png)

(1.37 MB 1024x1536 00171-1719150929.png)

>>16638 >>16639 I realized why this is happening now, it's because the reference preprocessors in ControlNet do their own attention hijack that interferes with SAG. Unfortunately that's kind of beyond the scope of what I know how to resolve.
>>16670 these people are going to get financially fucked in the coming years. have some compassion for them even if a lot of them are kinda dumb
>>16680 >have some compassion for them
>>16680 I have compassion for those who aren't twittertards on eternal moral crusade. They will adapt to new reality though. Twitterfags and redditors can eat shit and I hope they'll be homeless and miserable.
>>16680 Asking people here for compassion who like to shout nigger, trannies and jews left and right at everything is just not gonna work lol
>>16683 >who like to shout nigger, trannies and jews left and right at everything yes, and? what's your point, mr niggerfaggottrannyjew?
>>16684 u proving my point
>>16686 most artists are mentally ill, greedy and huff their own farts i call that "spiritually jewish behavior"
>>16687 oh... your initial response was so ridiculous i assumed you didn't actually mean what you were saying lol. moving along
(3.08 MB 1280x1920 catbox_y5flzc.png)

(2.87 MB 1280x1920 catbox_01ag8o.png)

(2.60 MB 1280x1920 catbox_te1y4z.png)

(2.59 MB 1280x1920 catbox_x62w4m.png)

Getting back from voice stuff to do some more loras. Oke Yazawa https://mega.nz/folder/2FlygZLC#ZsBLBv4Py3zLWHOejvF2EA
>>16689 >voice tech is the open source stuff getting better?
Does anyone know where this Kobayashi anime style is? This person's pixiv is full of pics using this style but I can't find it anywhere.
>>16690 Don't know about TTS but for voice conversion RVC is a little better than sovits or diff-svc. Training times are quick and pitch guidance is pretty nice. For making AI song covers you still need good source audio and Ultimate Vocal Remover will only get you so far. There's still no public tech for isolating specific voices unless you get lucky where you can just separate the center and stereo channels.
What are your favorite models right now? From what I've heard, Based64V3 > all other based mixes and AOM3 is generally not liked over AOM2. Does anyone still use AnythingV3 or AnythingV4? I also saw some interesting models by a someone on Civit. CyberAlchemist. The Loras they posted like "Detail tweaker" work exceptionally well, so I was wondering if anyone messed around with their Animix, Aniflatmix or Ambientmix models
(1.56 MB 1024x1536 catbox_p19kjz.png)

(1.57 MB 1024x1536 catbox_vphn6f.png)

(1.84 MB 1024x1536 catbox_p54gik.png)

(1.55 MB 1024x1536 catbox_k7jtdp.png)

I need ideas what to gen after nearly cooming myself to death
>>16693 NAI for most shit Based64v3 for some styles Custom furry mix for furry and/or futa
anyone have a guide for regional prompt
>>16688 ok kike
>>16691 >lolis with huge asses and thick thighs heaven
>Try out locon's through ez scripts >They feel really undercooked Did 32dim - 16 alpha 4k steps with a slower learning rate but thought they might fry honestly but they don't get character hair color consistent and seem weaker, play nicer with other loras and stuff from some testing.
fuck big corpos and elites, faggots are trying to restrict open source AI shit. Download every model and get ready to back it up and upload them through torrents
>>16700 >me being a data hoarder knowing DAMN WELL this would happen like clockwork tbh,fuck em spread the tapes.
>>16701 I'll know this shit is real once Civitai and Huggingface is told to take down all their models and code. I know we have underground backups but these power hungry faggots are probably going to do shit with hardware too if they don't get their way.
>>16692 >For making AI song covers you still need good source audio and Ultimate Vocal Remover will only get you so far. There's still no public tech for isolating specific voices unless you get lucky where you can just separate the center and stereo channels supposedly synthesizer V's AI banks are actually ai-trained and it works so well it might genuinely be nip magic
>>16702 >but these power hungry faggots are probably going to do shit with hardware too if they don't get their way. it would be so fucking funny (not really but you get what i'm saying) if the kikes burned nearly 3 decades of hardware backdoors just to prevent the average coomer from accessing AI image/text gen
>>16704 is there some piece of news i'm unaware of? all that happened so far is the patreon and pixiv shit right?
>>16705 Open AI Kike's CEO was trying to fear monger to congress to push for regulation and licensees for AI so they can finally have an advantage over it which google leaks showed that they were whining that they "have no moat" over open-source AI >But the uncomfortable truth is, we aren’t positioned to win this arms race and neither is OpenAI. While we’ve been squabbling, a third faction has been quietly eating our lunch. I’m talking, of course, about open source. Plainly put, they are lapping us. Things we consider “major open problems” are solved and in people’s hands today. fucking faggots just want monopoly over the market
I wonder if this stuff will get paywalled to hell and back in the future, everyone wants to profit one way or another from this somehow
>>16706 >fucking faggots just want monopoly over the market "OPEN Ai" lol, lmao didn't the kike ceo get super uppity at stock image companies and such? "lol you can't stop us from using your data, we'll find a way" kike hubris sure is something
so nothing happening and people just complaining for now?
>>16707 idk but we need like a contingency plan in case shit hits the fan and AI software gets restrictive by "the man" or something worse, I think /g/ has most of civitai backed up (not sure how much I just know they're backing up models) and I think they're on the right track despite their absolutely garbage /sdg/ general. Theres lots of money on this table and kikes are sure to not just leave it laying on there.
>>16710 can't wait for either intel or more likely nvidia to prevent you from starting any AI-related workload on consumer hardware, either by pushing stealth updates (through drivers or windows update or anything really) or by putting some microcode on newer gpus they might try harder this time unlike the whole LHR debacle
https://github.com/anonCantCode/sd-webui-send-to-controlnet Hacked this together (99% borrowed code) out of annoyance for myself earlier, but posting in case any of the anons here are interested. Not pretty but it functions; no idea why the ControlNet extension doesn't already have something like this. Loopback is only for the batch stuff, apparently?
Apparently the new update to the webui allows for offline usage, did anyone test it out here?
>>16654 I still can't reproduce that image no matter how hard I try to replicate the settings
(14.96 KB 493x150 anustek.png)

i might be able to bake soon™
(245.21 KB 384x640 Rinze.png)

>>16715 GL anon may you uoh with us when you get it updated.
>>16716 less about updating and more about building reports about ryzens exploding on ANUS boards started coming in the same day i wanted to start building
>>16715 god this shit absolutely toasts me. Thank god I couldn't afford an Asus mobo
>>16718 >god this shit absolutely toasts me it also toasts cpus :DDD
(1.18 MB 1024x1536 00093-2614442089.png)

(1.55 MB 1024x1536 00010-861570866.png)

(1.92 MB 1024x1536 00008-182143619.png)

(1.93 MB 1024x1536 00094-961103337.png)

(1.46 MB 1024x1536 00002-1986821808.png)

(1.63 MB 1024x1536 00007-2736027253.png)

(1.71 MB 1024x1536 00281-264260466.png)

(1.85 MB 1024x1536 00269-4078593046.png)

>>16720 zamn those are some THYUCC outlines
>>16722 i wish i knew more good artists that had thick lines like that because i really enjoy making loras of their work
(2.01 MB 1024x1536 catbox_i8fxn4.png)

(2.06 MB 1024x1536 catbox_rz7p94.png)

(2.02 MB 1024x1536 catbox_cwdyjc.png)

(2.04 MB 1024x1536 catbox_79iaec.png)

>>16720 Looks pretty nice, I'll give it a try later.
How would you prompt something like this, semi realistic clothing but still keeping the animeish style of the girl
>>16720 I like it, thanks
>>16726 Looks good with the 0.3 Zankuro mix as well and I will definitely use this more.
>notice negative getting long, so put an intentioned BREAK in it >for some reason doing that gave me more android girls Well, alright. I'm not complaining though.
(415.63 KB 512x680 catbox_vbamau.png)

(550.01 KB 512x768 catbox_vddgcn.png)

(408.82 KB 512x680 catbox_k910fd.png)

(432.01 KB 512x680 catbox_p5l8a6.png)

>>16693 Recently found vela-mix, I like how it looks and the body type it generates. Though I have some issues with some loras, where it seems to generate way too many features on the character, so I have to reduce the strength.
>>16725 sasuga cutoffs
>>16727 nice. catbox?
(1.50 MB 1024x1536 catbox_sbnzkx.png)

>>16729 nice thanks anon
>>16725 i would mix in a realistic model at low power
(976.21 KB 768x1152 00061-1854286833.png)

(922.34 KB 768x1152 00118-4083506465.png)

(648.23 KB 768x1152 00134-3584463360.png)

(790.28 KB 768x1152 00063-2035942635.png)

This art style is too cute for genning lewds with.
Ez scripts anon when I set the args list to lora, it's using the newer locon tech. Is that under lycoris style stuff or can I use the Conventional Lora recommended settings on blue leaf's github? I don't know what to train these at using this setting tbh since the few I've made so far seem undercooked or don't get characters hair colors right of all things. It happens enough in gens I feel something is wrong with my trainnig. Rest is 10/10 and I could just prompt in hair color but I haven't had this kind of issue with consistency before. Pic related is Iroha / Mutsuki from BA
>>16737 disagree but I respect your opinion
>>16738 do you mean when you set the algo in the network args to lora? if so, yeah it uses kohaku's implementation of LoCon. generally I don't actually suggest following kohaku's settings for things like LoCon because he suggests you always be a alpha 1, which just isn't a good idea if you want to have decent training times. despite that I can give you some insight based on what I know. LoCon do take longer to train, however not by much. usually I can get a good result within 2k steps. My usual go to when training LoCon is dim16 alpha 8, conv dim 8, conv alpha 1. that being said, I have a feeling that for styles you can get better results by doing something like dim 4 alpha 1, conv dim 16 conv alpha 8. if you are not setting conv anything then you are just training a normal lora via sd-scripts, like we always have. I feel like I'd need more info to be more specific about this, but I'll be around (as I'm just working on the UI right now) so don't hesitate to ask me questions.
>>16739 It feels wrong for some reason...
>>16736 Bonnie and Ruka Loras please
>>16724 it says the catbox for your 3rd image doesn't exist
Has anyone read this paper? https://arxiv.org/pdf/2305.08891.pdf
>>16740 Yeah alpha 1 feels weak but like you don't want to go above dim 32 on these right?
>>16748 >>16740 Making sure I'm not an idiot in following your directions I set self.network_args to lora to turn on locon, then set a conv_dim / alpha there too right? self.network_args = { 'algo': 'lora', # lora corresponds to locon, loha corresponds to loha, ia3 corresponds to ia3, and lokr corresponds to lokr 'conv_dim': '32', 'conv_alpha': '1' }
can anyone share a catbox image sample for alexi's gpqbao lora? i want an image to use as a preview for the lora in the webui as well as using it as a sample prompt that i can copy+paste when genning
>>16749 yeah you want to set them there as well, that being said if you use the popups it should ask you for the conv dim and alpha if you set locon. usually for me, I try to keep below dim32 on both linear and conv. I pretty much only use alpha = half dim, on both linear and conv.
>>16751 I'll try again without alpha as 1, just do a 32:16 then on them. I kinda like them because they seem way more receptive to change, prompts and other loras so if I can get them working right it'd be cool.
>>16751 Also realized this was at the bottom of my json. Do I need to turn this true and set 32:16 respectively or is this just standard?
>>16753 no, don't touch those, those were from when LoCon was new, I don't use them anymore but are still there for compatibility sake if you want I can give you the json file I have for my recent Arona + Plana LoCon I made
>>16754 Yeah if you don't mind just to compare settings.
>>16755 >>16754 And what the locon looks like. example images etc.
(512.14 KB 512x768 catbox_vvk7s7.png)

(423.42 KB 512x768 catbox_qhnizp.png)

(412.03 KB 512x768 catbox_7d7clu.png)

(516.49 KB 512x768 catbox_1dwfds.png)

>>16756 https://civitai.com/models/60885/arona-and-plana-or-blue-archive-or-character-locon examples there, json: https://files.catbox.moe/1ar7io.json i'm not able to gen right now, but here are some un-upscaled gens as well (slightly cherrypicked, but it really doesn't take much to gen them correctly)
>>16757 How's 1 epoch working out for you? Been usually doing multiple epochs and was gonna try your 3k steps but I think it'll cap out an epoch if it goes over steps right? So you want to pick one or the other.
>>16758 it was actually 8-10 or so epochs, steps doesn't eliminate epochs, it just sets the number of epochs based on the step count.
>>16759 >>16758 oh somehow I managed to ignore the second part of that sentence. the way sd-scripts handles steps is it will continue to go until the exact step count specified, that means it will stop no matter where it is. Even if it's in the middle of an epoch. it will end right there and save the final epoch. that does mean that the final epoch can be trained towards a bias because it isn't a full epoch, but if it is you can just go back one epoch. it means that I don't have to really calculate the epochs, I generally try to keep my image count within a reasonable 100-300 range though, so usually I get anywhere from 8 to 16 epochs (depending on if I'm on the low or high end)
>>16656 >>16654 >>16659 anyone have a good catbox example image of fisticuffs to show what it can do?
>>16760 So if I set it to 3k steps how does it decide to end an epoch, or where since your json tells me there's 1 epoch.
>>16763 sd-scripts handles it for me, it knows how many steps are an epoch, so it knows how many epochs to do. the reason why the json says 1 is because that's the default value. the steps value takes precedence though. This is actually how it works in sd-scripts, so in order to get epochs to work I have to ensure that steps isn't set.
(2.60 MB 1280x1920 catbox_9ohkiz.png)

(2.82 MB 1280x1920 catbox_rzrl8o.png)

(2.82 MB 1280x1920 catbox_d0o4ti.png)

>>16762 Quick comparison also with another lora I will be uploading shortly. You can kinda see the reason I wasn't originally too happy with it because it forces character details but if that's what you like then all the power to you.
>>16764 Is that a default thing in sd-scripts or something to be configured .
(2.17 MB 1024x1536 00588-181300678-grin.png)

(2.11 MB 1024x1536 00043-3288607810-evil grin.png)

Yes officer, I'm guilty of genning cunny.
>>16766 Yes, as far as I remember. I know that I don't do anything special for calculating epochs when steps are given, I just give the steps count to sd-scripts and they handle it.
>>16768 Yes as in a default thing
4060ti 16gb might be a very interesting "budget" gpu! I'd hoped for it to be a bit cheaper than 500, but if it actually sells for MSRP that would be pretty cool
>>16770 five hundred dollars isn't budget. stop doing nvidia's jewish wheedling for them.
>>16770 >>16771 not only it's not a budget option but the extra 8GB cost $20 to add It's a genius way to sell both the regular 8gb and fuck people over or get people to realize that the 16gb is shit value and "might as well get a 4070 at this point"
what is this kike doing
>32dim LORA gets more consistent details than 128 and 64 this must be some RNG shit, it's why baking can fuck up your perspective sometimes
(6.79 MB 2560x1882 catbox_hy7pq5.png)

(5.30 MB 2560x1882 catbox_f1ubn6.png)

Hmm not really happy with either but I already uploaded the "fanbox" one. Main difference is the pixiv one was trained with more variety on 1 repeat, approx 17 epochs. "fanbox" was trained with a lot of the same images (with small variances) but balanced repeats.
(5.60 MB 2560x1882 catbox_zpyswt.png)

>>16776 On second thought the pixiv one is better because it doesn't just completely ignore the prompt
>>16775 isnt this a case of different alpha or LR
(5.53 MB 2560x1882 catbox_fh0j5r.png)

(5.49 MB 2560x1882 catbox_uwaxgo.png)

(5.53 MB 2560x1882 catbox_dxdfvo.png)

(5.50 MB 2560x1882 catbox_epe3eq.png)

so no one is posting useful LORA info anymore on 4chan hdg, for fucks sake it's so long gone and this thread is so dead that it takes days or weeks till someone posts useful info
>>16780 you sound like you just hoard shit and never gen or train relax buddy, no one's chasing you
>>16780 how often do you need a new post about a mild refinement of the same old techniques
>>16782 There's a lot of new updates to sd-scripts that can add factors to the final results, probably not as huge as the tired arguments of repeats, epochs, and learning rates but it's still something worth looking into
>>16769 >>16768 Might just revert this and stick with loras for now, I can't seem to finagle this to give me a consistent chara's traits without hard prompting them in, 5e-5 might be too slow of a learning rate for locon's I guess.
>mfs putting "Prompt Engineer" in their twitter bios deep and disgusted sigh
>>16785 at least it's gonna be shortlived, there's no way it doesn't become possible to finetune LLMs to output image prompts even if multimodal models remain huge and shit
>>16767 catbox for that first one?
(4.20 MB 2048x2048 LOCKER ROOM BUTT LEAK.png)

(4.28 MB 2048x2048 LOOK WHAT YOU DID.png)

(4.03 MB 2048x2048 PAINAL 1.png)

Been at it less than a week. Still feel like a noob but getting it's getting easier. I'm having trouble getting the sharpness that I'm seeing in examples uploaded to civitai and I'm not entirely sure why. Is it the hires fix? I run out of vram when I try to turn that on even though I seem to have plenty for everything else.
>>16788 How are you getting your renders to 2048x2048 if you're not using hires fix? I personally start at 512, hires to 1024, adetailer inpaint faces/eyes, manually inpaint any other problem areas (pussies especially), then SD Ultimate Upscale + ControlNet Tile to 2048+ and that gets me results I find to be reasonably sharp. If you're trying to upscale again past 2048 without using tiles, that would certainly be why you're running out of VRAM.
>>16789 I gen, img2img/inpaint at 512x512 then I run the 4x upscaler. I can hires a 512x512 but if I try to run a 512x768 portrait with hiresfix I get the VRAM error. I'm not really sure that I even need to run hiresfix at all. I'm just trying to figure out how to make my stuff look more better like examples I see
(4.76 MB 2048x2048 WET HERSELF ON A DATE.png)

(174.27 KB 1024x1536 CITRINEDREAMEXAMPLE.jpeg)

>>16789 See it's like they end up decent enough but for mine (left) when you zoom in especially on the face there's like this acryllic painting type effect I keep getting. It's not necessarily bad in itself but I'd like to be able to do it when I want and not when I don't y'know?
>>16792 Scaling up to 2048x3072 without a tiler (as hiresfix/the normal upscaling algo does) requires 50% more VRAM than 2048x2048 does. I don't have the numbers, but I don't think any consumer hardware has the VRAM to support that. Like I said, I suggest to upscale to 2x only, inpaint, and then use tiled diffusion (ControlNet or otherwise) to upscale to 2048+.
>>16793 Stylistically I wasn't too in love with these, so they got left at the 2x stage, but the sharpness is there.
>>16791 Hiresfix is why examples you see look better. If you can't do 2x try 1.5x, or try different resolutions. If you have live previews enabled set it to Approx NN or turn it off to free up vram. For fererence I have an 8GB card and can hiresfix 2x with 4x-UltraSharp upscaler from base 640x960 res.
(2.01 MB 1024x1536 catbox_h6msuq.png)

(2.02 MB 1024x1536 catbox_129sdu.png)

>>16793 Ok cool. I actually just discovered that ControlNet existed yesterday and downloaded the models. Haven't tried to use it yet, slightly intimidated. >>16795 Nice tip thx (I kinda like watching the preview actually but it's not like it really shows that much). So doing Hiresfix on the initial gen produces better results than working only in 512x512 and then upscaling? I like how more quickly the workflow goes when I stick to the small size until I'm done but ultimately results are what's most important.
>>16793 Sorry if this is dumb but to make sure I understand. You gen in 512x512, then when you have something good you upscale 2x, continue to work on it, then use ControlNet to 2x it again? Tiled diffusion just means "use ControlNet" right? Which ControlNet model do you use (like depth vs lines vs whatever they are)? >>16794 hmm I think I'm starting to get it. So you gen'd those at 512x768 with 2x hiresfix at initial gen? Or you gen at 512x768 then 2x one you want to work on with upscaler, inpaint at 2x, then 2x again when you're finished? As an aside, does anyone have an opinion on the AddDetail LORA? That seemed promising but in practice it just like multiplied that acrylic painting effect I was getting by 10x (see example). NGL, I kind of like how hilarious this one turned out but it's not what I was trying to do. One of my earliest ones before I figured out how to get rid of that goddamn steam. What is the deal with the steam
>>16794 P.S. What model/vae were you using there? I've been trying out a lot of them and struggling to generate petite frames that don't have rippling muscles and/or pregnant stomachs like that gonzo hentai style I can't stand where every face is an ahegao.
(678.31 KB 768x1152 00009-1849526644.png)

>>16803 naisu
Oh boy, here comes a text wall. >>16797 Img2img upscaling is literally the exact same process as hiresfix so long as all the settings are identical. >>16801 I gen in 512x768.When I get a seed with the overall look I'm going for, I upscale normally to 2x -- 1024x1536. I use a custom ESRGAN model, 4x_fatal_Anime_500000_G, which I feel gives me better results than the built-in stuff. Once it's at 1024x1536, I inpaint any details I find lacking -- faces are run automatically with adetailer at upscaling time since I find those almost always look somewhat lifeless out of the box, hands, genitals, etc. Inpainting at 512x is practically useless for detail, there just aren't enough pixels. It's easier to get big blocks (extra limbs) taken care of at the smaller scale, though, if that's needed. Once everything looks good at 1024x1536, and I'm happy with the style, I take that image, and upscale it 2x again using the Ultimate SD Upscaler script/extension and ControlNet 1.1's tile model/tile resample preprocessor. This is what I mean by tiled diffusion. It's internally separating the image out into tiles so that it isn't having to work on the entire thing all at once (and running out of VRAM as a result). There's another extension for a1111 that did tiled diffusion separately from CN, which is why I'm saying it like that, because that's another option (although I think CN's implementation is just better at this point). The other CN models do a variety of other things and aren't necessarily useful for upscaling in this way. Things like composing the image based on a depth map, or attempting to match a pose. Examples of each are on their GitHub. https://github.com/lllyasviel/ControlNet-v1-1-nightly As for model, I'm using Based64/66 (each image in 16794 was generated with one, see filenames) with ClearVAE. I haven't run into the situation where the detail off AddDetail is actually something I'm looking for, I tend to dislike the super sharp look, but it's in my style randomizer wildcard for when I'm feeling lucky.
>>16805 you can also get a similar result as using Ultimate SD upscaler + controlnet tile if you use multidiffusion + controlnet tile, it's what I usually use.
(756.34 KB 1024x1536 00028-3323809903.png)

>>16804 Thank u >>16805 >Img2img upscaling is literally the exact same process as hiresfix so long as all the settings are identical. It is if you tick "With img2img, do exactly the amount of steps the slider specifies (normally you'd do less with less denoising)." setting Dunno if it's present on newer commits though
>>16807 Never knew about that setting, despite using img2img to upscale for a whiiile now. Do you know how it impacts the steps and denoise? Does it simply halve both?
>>16808 With this setting it actually works similar to hires fix. By default, it reduces the amount of steps according to your denoise. If you chose 20 steps with 0.5 denoise you'd actually get 10 by default. Check that in console when you do img2img. With 20 steps 0.1 denoise it would actually do 2 steps. With the setting enabled it will actually do what it says. I am on a9fed7c3 though, so it might have been changed on older commits.
(968.67 KB 1024x1536 00024-480112492ed.png)

>>16809 >so it might have been changed on older commits I meant on later commits obviously
>>16807 It's still there on Vlad's fork. If Voidy got rid of it, unless he just made it always-on, he's more of a damn fool than I thought. IMO there's no reason that setting should ever be turned off (and yet, it's off by default). If you need to do less steps, you can just.. do less steps.
(1.35 MB 1024x1536 00031-4019622300_cleanup.png)

>>16811 >IMO there's no reason that setting should ever be turned off (and yet, it's off by default). If you need to do less steps, you can just.. do less steps Certainly agree. Anons frequently said that img2img upscaling was inferior to hires fix and I suspect this default behavior of reducing steps is one of the main reasons.
>>16810 Thanks for the info! It seems I have been upscaling at only 8 steps, and I've still been pretty happy with the results. I'm now able to actually reproduce the normal hires fix results in img2img, so thanks for that! I've changed the settings because obviously this halving / changing of your selected steps should have never even been a thing. also, catbox for this image?
>>16813 Yeah it generally seems like not much steps are needed for decent-looking upscales, but it's still nice to have more control over the process. You're welcome anon >catbox https://files.catbox.moe/4bujnb.png https://files.catbox.moe/whd520.png
>>16805 Awesome, I appreciate the thorough reply. It sounds like I need to start learning ControlNet now. Today I tried out generating at 512x768 + 2x hiresfix and started getting much better results right out of the box. But then the workflow slowed to a crawl as I only had the VRAM to gen a single img2img copy per run. If I understand correctly, though, CN will solve that. I agree that 4x_fatal_Anime_500000_G seems to work better than the AnimeDB most guides appear to tell you to use. My upscales with the latter all come out flat. What's that style box about anyway? I've seen models categorized as "Styles" but they all seem to be LORAs so I've had no idea what to do with that style dropdown in the UI. I think just switching to a portrait layout may have helped me out a lot, as I did this test image 512x768, upscaled it 2x. Then I got impatient trying to img2img it in batches of 1 at the 2x resolution, went back to using the 512x768 in batches of 4 then upscaled. For what it's worth, I compared the end result of both and they seem to be essentially the same unless I'm missing something. I named her Tanuki Maria lol
>>16814 >provides both the txt2img and the img2img seed You are valuable one, anon >>16815 As far as I know, the style box dropdown is for saving and loading text prompts. Style Loras are there to change the overall style of the output. Like used here for example to replicate an artists style >>16737
>>16816 ohhhh you right. I'd been saving them in notepad like a moron
>>16817 Honestly, I prefer that to a1111 styles anyway -- I have all of mine saved as individual text files for the Dynamic Prompts extension, which lets me keep them categorized, have more intuitive control over how they're ordered in the prompt, makes them easier to edit since they're just plaintext TXT, and makes them easier to use in other workflows because.. they're just plaintext TXT and not an asinine CSV.
>>16818 I'll have to look into that extension. I haven't really looked at what's available at all, other than installing the extension for LYCORIS because I needed it to use them. Probably some other quality of life stuff hiding in there
>>16817 >>16818 There's an extension that's like the built in LoRA browser with image previews for styles, but it uses its own file, so you'll have to define the styles again. https://github.com/eideehi/sd-webui-better-styles It also allows to save other parameters like steps, cfg scale etc., or at least that's what the gui shows, but it never worked for me. I haven't updated in some time, so maybe it's fixed.
Lora chads I need help, I got a character that has a one side up hairstyle but some gens end up with two side up, what should I do?
>>16821 You mean like if he has two horns when he's supposed to have one? For things like that I open in Krita, copy/paste a section of the background over the one I want to get rid of then run the result through img2img with low denoise in batches of 4 until my butchery gets blended out. I can't say it's the best way but it's what I do. You could also use inpaint to reroll just one side of their hair until it turns out how you want.
>>16824 was wondering if there was something with the text encoder LR that can fix it because I'm already reaching 1e-4 with the unet
>>16825 You could try playing with the weighting but if I had to guess it'll probably just have some RNG attached to it. I just started myself but so far LORA have been a crap shoot for me. Sometimes they fuck up more things than they help.
>>16765 why not just use tiling upscale and skip hires fix altogether?
>>16805 dumb question, but the 512x768 is no different quality wise than 768x512, right? it just changes what content can fit (portrait vs landscape)?
>>16820 can't we just get this functionality by saving generated images into the image browser extension folder and then sending the picture to txt2img or img2img via said image browser extension tab or am i missing something?
>>16817 personally, i just keep a folder of pictures with styles i like and rely on the embedded metadata, and also a search program annoyingly called "everything" for when i accidentally save an inpainted version with a truncated prompt and have to go dig for the original
>>16830 catbox? o.o
maybe i'm a schizo but i feel like the further we mix shit the worse the model gets at generating built-in characters
>>16833 the more children between you and the original models is definitely going to change things. i don't think it's necessarily for the worse if it's a competent mix, but I'm not an ML scientist either
>>16833 Not schizo at all. Even negs affect how well the model gens the NAI-known characters. Without negs it "knows" the characters characteristics fairly well, but once your ramp worst quality, low quality to 1.2+ the resemblance starts to decrease
>>16836 >but once your ramp worst quality, low quality to 1.2+ the resemblance starts to decrease MOTHERFUCKER IS THAT WHAT'S BEEN FUCKING ME UP? I'VE BEEN USING 7TH LAYER NEGS ALL THIS TIME FUCKING FUCK I KNEW IT
>>16833 yes tag dementia is a known thing its in one of the guides so its not schizo at all
(3.03 MB 1536x2304 catbox_r5xh0l.png)

I'm at the point where i barely even go on boorus during masturbatory sessions anymore. I just fire up Stable Diffusion and get to work
>>16836 What characters does NAI know anyway?
>>16840 basically anyone with more than 1.5k pictures on danbooru
>>16840 all the characters >>16805 i haven't used it yet, but from what i read, isn't controlnet's tile resampler and tile model meant to be used standalone? seems weird that it would work with ultimate SD upscaler unless i'm misunderstanding
>>16840 https://seesaawiki.jp/nai_ch/d/%c8%c7%b8%a2%a5%ad%a5%e3%a5%e9%ba%c6%b8%bd/%a4%a2%b9%d4 Japs have autistically charted what characters it knows (most of them anyways), left hand bar for navigation
>>16837 Yeah I think 1.4 is honestly an overkill generally. 1.2 works better and has more sovl most of the time. Subjective, of course
(973.84 KB 1024x1536 00016-3882055313.png)

(1.22 MB 1024x1536 00009-1989624664.png)

(1.99 MB 1024x1536 00013-3020566809.png)

(2.53 MB 1280x1920 00019-4172888044.png)

(1.46 MB 1024x1536 00104-643157593.png)

(1.71 MB 1024x1536 00016-486486896.png)

(1.75 MB 1024x1536 00044-989009030.png)

(2.26 MB 1024x1536 00204-2499085938.png)

>>16846 >>16847 Thanks. Do you just lurk boorus to find artists you like or you already have some kind of prepared list you already knewe and wanted a lora of?
>>16848 i have a list of artists i want a lora of. i also have datasets to clean up of artists i want a lora of. i then ignore both of those things, and sporadically find some random artist on a booru, then make a lora of that instead.
>>16852 damn I wish I had your libido maidanon
Scraping EijiERJ just in case someone figures out a way to get GOOD feet
>>16829 I guess, haven't thought about that. Though the thing about styles is that you can select multiple, while sending to txt2img or img2img overwrites the previous one, so it's good to make a bunch of styles that focus only on a part of the prompt, like one that's just quality buzzwords at the beginning, another that's a specific background or setting, or having different negatives like one for the usual (worst quality, low quality:1.4), another for all the negative embeds, and another one that works only for a specific character, style etc., which you can freely combine without having to fish for a specific image, deleting most of the prompt then loading another image to copy paste another part and so on.
>>16853 Lifting, ashwagandha and zinc maxing.
>>16812 cute!
>>16842 It ultimately ends up that it's using CN tiling for the initial upscaling, then redrawing SDUU tiles. Might be a bit moot, I guess, I haven't done a lot of testing of either of them alone, but it doesn't seem to hurt anything, anyway.
>>16847 catbox for #2 good sir plzty
>>16856 >>16853 my issues not the libido so much as hard flaccid or w.e tf it was called. ssri is probably partly responsible
>>16860 Not a doctor but I am a pharmacology understander who successfully treated my own depression after years of doctors and therapists not helping shit. SSRIs are poison. They make your brain release more serotonin but they don't help it make more of it, so they actually make things worse by literally running your brain's serotonin stockpiles down to zero. At that the only thing it's doing is keeping SSRI withdrawal syndrome at bay. Doctors don't understand this so they just increase your dosage or add additional drugs. Order some L-Tyrosine and L-Tryptophan (5HTP works too and can be easier to find than the latter) supplements. Those provide your body with the ingredients to make as much Serotonin, Dopamine, and Norepinephrine as it needs, If you're taking those then you can quit cold turkey with no withdrawal symptoms.
>>16827 You could, but it's not really needed when going to, say, 1024x1536, and I would still suggest taking the time to scale to 1024x and do a pass of inpainting before upscaling further so it has to do less guessing about what you want with the eight times as many pixels (vs 512x) at 2048x+. >>16828 For the most part. Depends on model/LoRA, as some are trained in portrait/landscape rather than square, and so can act significantly differently between the two.
>>16861 >They make your brain release more serotonin but they don't help it make more of it I'm another armchair pharmacist and I do agree that SSRIs are poison but it's not a serotonin releaser, it's a reuptake inhibitor which is a different thing. But it's irrelevant anyway imo, trying to actually understand this stuff is almost as useless as understanding lora training parameters. For someone without deep knowledge of the topic it's just observation, and SSRIs do indeed seem to be fucking poison for 99% of population. The only antidepressant drug class which seems ~kinda useful is MAOI which are kinda dangerous for other reasons.
>>16856 I guess I should finally start lifting then. Ashwagandha makes me way too calm and I probably supplemented enough zinc before
>>16846 Cool I like this one
>>16860 the depression -> SSRI -> erectile dysfunction -> sexual frustration -> anal masturbation -> anime transgenderism pipeline, very sad to see
>>16867 Fuck off pol tourist
>>16864 You are correct. I am aware that they actually create more serotonin in the active space by inhibiting reuptake rather than by stimulating release. But I was afraid if I got pedantic on that I would triple the size of my post and it wouldn't make the person reading it understand what I was trying to get across any better. Your brain takes the serotonin back up and reuses it but the SSRIs block that process making them spend more time in the active space. But if they're not being taken back up then they're not ready to be released when they're needed either and eventually you run out of reserves. The main idea I was trying to get across is the same though. SSRI = run out of Serotonin. Solution = stop taking them and take supplements to help your body replenish your reserves. MAOIs don't really have much if any more efficacy than the other classes of antidepressants and they carry a lot more side effects and risk. The best antidepressant is a combination of good nutrition, sleep, exercise habits and a well-centered mind. If that's not enough then Psilocybin, LSD, and/or Ketamine are all actually effective at relieving depression symptoms. It just works differently where you take them once and you feel better for a while until you need to take them again and recharge rather than something you need to take all the time.
>>16860 I've always wondered how having your dick quit working is supposed to help you feel less depressed. Years ago when I was taking Celexa, I could maybe manage to jerk off about once a month, really just to make sure it still worked. Then my doctor added Risperdal and it started actually hurting when I came. Hooray
Fucking Civitai faggots ruining my plans to do an all in one LORA for an anime I was watching recently, good thing it was overfried though I just want to try my hand at the anime screenshot collecting shit with ffmpeg
>>16859 they're all in the previews folder for the lora
>>16867 did the risp make your dick harder or something? asking for a friend
>>16875 No it made my dick even less able to get hard then when I cum only dust shot out and it hurt.
>>16875 >risp chinese girl with a speech impediment?
>>16867 ED anon here: what's the cannon explanation for why any given doctor or psychiatrist prefers SSRIs over supplements? this is excluding the shitty doctors that only care about pharma handouts. as for pharma, i can picture them going the SSRI route due to abundant supplements, leading to a race to the bottom on prices
>>16878 >>16870 sent that previous reply to the wrong anon
>>16876 ah. i asked just because i've had erections that were too hard in the past which made coom painful
>>16878 The medical field is captured by moneyed interests in the same way as the government. Lots of doctors' offices when posting job listings will even list as a perk of the job that you get tons of free lunches because pharma sales reps will just roll in at lunch time with surf and turf for the entire office. Also M.D.s (which includes psychiatrists) actually get surprisingly little training with regard to drugs and pharmacology, for them it's more about diagnosing the problem and then following a treatment schedule. The only people who really get taught anything about the how/why of drugs on a fundamental level mostly end up working in labs trying to make new ones rather than treating people.
>>16880 Yeah I can't really speak to that but if you want your dick to be less hard then antidepressants are the way to go. They're such reliable dick killers that they actually prescribe them off-label to treat premature ejaculation.
>>16878 Also yeah, there's not much profit to be made in telling someone that they can fix their crippling depression by buying some shit that's not even drugs, literally just nutrients, and a year's supply sells on Amazon for like $50. So exactly whose incentives are aligned in a way to drive them to encourage this? Whereas when you put someone on SSRIs, you pretty much ensure you have a paying customer forever because they will stay depressed (and get worse) as they fully run out of neurotransmitters. But they'll keep buying them and even believe that they help because of how terrible they feel if they try to stop taking them. But that's the exact same thing as saying that Heroin is helping you because you feel bad if you don't take it.
>>16868 lol is he wrong tho?
>>16868 Found the tranny. >>16867 is absolutely right. If you're depressed alcohol is still less poisonous than SSRIs and you get to keep your penis too. TTD!
>>16886 >>16888 Stop bringing up troons and jews out of nowhere, I see enough of this garbage on cuckchan, can you just post AI-generated girls without identity politics just for fucking once?
>>16889 >NOOOOOO DON'T GATEKEEP Found the subversive kike! Wait, that's redundant.
>>16889 I don't see you posting gens either, kike.
>>16886 Ive been on-and off anti-depressants for ten years and I have never wanted to shove anything up my ass, now shut the fuck up about this pol shit and post some gens or go back to your containment board on cuckchan
>>16892 >bring up globohomo once >absolutely butthurt Curious!
(1.36 MB 1024x1280 00048-474276950.png)

>>16892 They just can't stop obsessing over ugly trannies and jews it seems they can't fucking help themselves. It's just tiresome.
>>16892 >>16894 samefagging huh
>>16894 i enjoyed making this post but things got very gay and tedious very quickly because of it. i'm sorry.
>>16896 never apologize for being based
>>16896 A bit of bait is fun but as you can see this theme summons actual schizophrenics and fucks up the whole thread
>>16898 >summons actual schizos rent-free
>>16889 Your reaction to what were clearly sarcastic jokes says a whole lot more about you than it does about anyone else, buddy.
>>16892 Dude it's ok. Shove all the stuff up your ass that you want. No one cares but you. Your rage is misplaced.
>>16899 >>16900 >>16901 Thanks for your input, you could just type it all in a single post buddy
>>16896 It was a good post. They will either calm down eventually or we will keep poking them and that will be fun too but either way it was funny so who cares
>>16902 I could've if I had about 75% less ADHD than I actually do. Sorry if I made you jump around
>>16900 not him, but damn you're retarded fucker, go back to pol brainlet
>>16905 where did pol touch you?
Told you, we need to gatekeep this place from the 4feds troons and kikes. The relentless screeching about /pol/ is extremely funny though, they wouldn't have lasted a single day on old, true /pol/.
>noooo you MUST obsess over disfigured troons and filthy kikes 24/7 or you don't BELONG to my sikrit club!!!
>>16908 >sikrit club!!! Thanks for outing yourself, outsider. The only fags who call it that are rejects from 4feds. Bravo!
I'm "reading" Secret Journey again after years and years, I forgot how good it is. I'll make a po-ju LoRA just in case we get a better SS/oneeshota LoRA in the future.
>>16907 look guys im so special!!
>>16909 NTA but what do you think this place is and where did all these posters come from? >all this shitposting wowe. Kinda funny that so many of you ousted yourself as mentally ill taking medication, you could kinda tell with people like maidanon but its always nice to have a confirmation schizos are actually schizos.
>>16912 what's wrong with maidanon...
>>16912 >NTA but what do you think this place is and where did all these posters come from? >he then proceeds to call maidanon a schizo This kike can't even act subversively without giving himself away one way or another. Give up.
>>16914 just stop posting polshit, you have your own containments to circlejerk, why ruin this place?
>>16912 No sane/happy person will visit any kind of imageboard.
>>16915 Again, where did /pol/ touch you? Why does it live rent-free in your head? Is your brain so fried and are you so terminally online that you automatically associate this kind of talk with /pol/?
>>16917 where did you trannies touch u? or jews? u do love talking about them, rent free huh
>>16917 Stop playing stupid, you are the exact fucktard who frequents pol 24/7 and brings up jews and trannies with every fucking discussion every where on the internet. You are equally cancerous as the twitter lefties, same kind of idpol garbage. Kill yourself freak.
>>16914 no I genuinely only scroll down this place every now and then, I post in the main thread so im not involved in all of this. Dunno wdym with calling maidanon a schizo - its not some new take that there's something weird with him. Not that I particularly mind since he's one of the nicer ones so maybe I shouldnt have made it sound so aggressive. >>16916 post election take. Sadly. Used to be different
>>16918 My ex best friend from high school trooned out, kikes refuse to repair potholes and that costed me a tire and scratched up my FK8's splitter. >>16919 By the way, if you're trying to use someone's argument against him, try doing it BEFORE you act subversive for the entire reply chain/BEFORE you out yourself as a kike. Not after. We can see what you're doing.
>>16921 if thats the reason hes not ur friend anymore damn you're one shitty friend then what can I say
I hate polniggers so fucking much bros.
>>16921 At least that guy avoided being friends with a faggot like you.
>>16923 >>16925 >he may be a tranny now but at least he's not friends with you! Why would you out yourself as a tranny?
>>16926 >this level of thinking oh boy I dont even know what to say, just hang in there buddy Im just glad ur not his friend anymore
>>16927 >this level of thinking This tactic doesn't work on me. You can try to pull off what you think is/would be a witty reversal all you want, it's not gonna work.
(1.37 MB 1024x1280 00012-3813358101.png)

Just ignore the polshitter and post AI girls.
>>16894 Kuro Lora?
>>16931 Cute belly
>>16933 *Slightly* pregnant.
>>16894 Please share catbox and Chloe lora
(1.22 MB 1024x1280 00055-474276956.png)

(1.17 MB 1024x1536 00022-2796207568.png)

(1.18 MB 1024x1280 00051-474276953.png)

(1.53 MB 1024x1536 00000-3345533198.png)

>>16936 Bless you sir! Have you, by any chance, stumbled upon decent Illya and Miyu loras (besides DGB repo)?
>>16937 Illya is promptable, dunno about Miyu. There's kaleid lora on cividdit that has Miyu in it, but it's style heavy
For all the hatred Civitai gets here, is there a viable alternative? Huggingface is hot garbage for discovery, and perusing our threads only goes so far....
>>16939 I do save plenty of loras from civit personally, but it's pretty disappointing to see a character/style lora that might be useful, but is actually another 100 repeat trash.
>>16940 Same for me, I save a bunch of lora from civitai, but I usually just do a quick judge of it before I download it based on preview images. Generally, if it's higher than dim32 (if it's a lora) or dim32, conv dim16 (if locon), I don't download it. Unless I'm very interested in it. There are some styles for example that looked really good but were dim128, so I would download those and resize it down. Fro resizing does a good job at reducing the dim (thus removing some overfitting/artifacts) without removing the style
Friend is getting this error, what do?
>>16943 install python 3.10 instead of 2.7...
>>16944 She did, same error. Do I just tell her to uninstall 2.7?
>>16945 The python in the PATH should be 3.10.9 one Change environment variable (Type "Environment variables" in windows menu) and check if the python in path is set to 3.10 probably easier to uninstall 2.7 unless you really need it
>>16946 Yeah I'm getting her through it, she enabled the add to path option in the 3.10 installer but webui-user still used 2.7 for some reason
>>16905 Literally never been to /pol/ in my life. I guess I'm just a 0.01% well-adjusted person, sorry if that triggers you. Maybe you'll feel better after you go scream into a pillow for a while about how the Jews are why you've never had sex. I posted like fifteen images I made yesterday. How many did you post? Oh, zero? That's what I thought.
>>16949 I'm trans btw if that matters
RAHHH if I stop screaming about trannies and jews for five seconds then those urges to put on a dress and squat on a horse dildo come back but I'm a real man so I know better than to give in to temptation RAHH there it is again TRANNIES! JEWS! RHARADSRHASRH
These sure turned out well straight out of txt2img imo
It seems like on Civitai, every example is run at CFG 6-8 for only about 20-25 steps. I'm increasingly wondering whether that's even close to optimal. For example, I spent a long time yesterday trying to get her top right in this image. I was in the process of trying to paste her left tit onto the right and use img2img to blend it back in when I had to go do something so I went ahead and just ran the prompt again, turning CFG up to 11 and running it for 300 steps. Came back and it had spit out the exact same image I'd been trying to make only perfect out of the box. Something to think about.
>uploading lowres images
>>16950 Is that why you're so angry then? That was more about the repressed anger I see coming out of people on here and how it usually stems from an inability to accept something about themselves that turns into irrational and unnecessary rage pointed outward at others. I've got nothing but love for trans people in general.
>>16954 Ain't got time for that shit. If I stopped to wait to upscale every picture I finished, make about 80% fewer in a day. If you like it then upscale it for yourself. If you don't like the image then why do you care anyway?
>>16953 Higher CFG scale increases creativity, but increases the likelihood of overcooking the image. I tend to adjust between a range of 7-11. What the sampler step count affects is going to depend on the sampler (duh), but I don't think any of them benefit much, if at all, beyond 50. I trend towards a maximum of ~40 with DPM2++, and ~25 with Euler A. Basically, you got that result from the CFG scale, and you just wasted time with that step count.
>>16953 >300 steps Typo? But anyway, I personally use 4-9 CFG. Higher looks too cooked for my taste. 4 helps salvaging some shit loras and looks pretty clean tbh. But the background are kinda fucked at low cfgs from what I've noticed so it's mostly good for simple
>>16903 >>16953 catbox? >>16940 i haven't trained a lora in my life, so this might sound retarded, but could there theoretically be a low effort method in the next couple of years that would unfuck a lot of the civit loras either by an easy to use retraining program or by a model to counterbalance the artifacts? i guess that's kind of what embeddings do somewhat, but they fuck with the possible outputs too much imo >>16942 eli5: how does reducing DIM size remove some artifacts/overfitting without removing the same ratio of style? are you saying large DIMs have a habit of retaining higher percentage of artifacts/overfit bs versus style? why would that be the case in first place if so
I think skipping out on LORAs that are above 32 dim is a bit harsh, 64 dim is my limit
has anyone added this to their webui? I find it to look fried compared to DPM test samplers https://gist.github.com/crowsonkb/3ed16fba35c73ece7cf4b9a2095f2b78
>>16961 I don't even have lora training installed locally so I can't convert dims. I have fuckloads of 128+ dim loras
>>16953 generally i use 4.5 CFG. sometimes 10+ is good but it almost always seems to fry a picture. can't find any particular rhyme or reason, i assume it's a lora-by-lora thing.
Anyone have a dragon quest lora for tabatha or the horse monsters?
>>16962 It's implemented in the dev branch, kinda shitty unless you double the sample steps and use the karras variant
>>16966 it just looks fried to me, I like the dpm test models that was posted awhile back because it actually takes off some of the orange contrast tint that klf8 VAE has
(2.61 MB 1464x2048 catbox_d9vqkr.png)

(4.02 MB 1576x2048 catbox_iu1no9.png)

(2.69 MB 1464x2048 catbox_cey49r.png)

(3.56 MB 1464x2048 catbox_lpnx7j.png)

>>16968 did you end up releasing this soft-3d model?
>>16969 Its pretty shit rn so i'm not sure what to do exactly. the base model for merging is kinda fucked due to SuperMerger. untill i can fix the base model to remove realistic nose and not break the coherency I won't be realeasing it
>>16970 gonna be frank with you, i don't mind the realistic nose
>>16971 it's not even realistic, it's just as109
>>16971 Its just a me problem the nose just needs to go. without breaking the coherency ofc if only the supermerger wasn't borked and the author was easy to talk to
I forgot what was the best learning rate for 1/2 alpha compared to network rank when it came to adamw8bit?
hey all, THE UI IS FUCKING DONE https://github.com/derrian-distro/LoRA_Easy_Training_Scripts that is all
>>16976 ok, maybe that is too short, make sure to read the readme, this ended up being a full re-write, and it doesn't have everything quite yet. queues and block weight training are excluded for now, until I figure out how I want to set them up. the readme also has a bunch of gifs showing how the UI works!
>>16953 wish it worked on my 6gb card
>>16976 I assume this includes all the updates Kohya has had since you last updated the popup version of your scripts?
>>16979 yes, it supports has the dadapt updates and the pyramid noise updates. I didn't include the dim_from_weights option because... I kinda forgot it existed, but sd-scripts is up to date.
>>16976 getting some weird ass error when trying to install >LoRA_Easy_Training_Scripts\installer\installer.py", line 163, in <module> main() File "LoRA_Easy_Training_Scripts\installer\installer.py", line 49, in main subprocess.check_call(f"{os.path.join('installables', 'change_execution_policy.bat')}") File "c:\python310\lib\subprocess.py", line 369, in check_call raise CalledProcessError(retcode, cmd) subprocess.CalledProcessError: Command 'installables\change_execution_policy.bat' returned non-zero exit status 1.
>>16981 odd, the execution policy bat should work fine, I tested the installer before I released it. how did you launch the installer?
>>16982 as explained on the github page, once it didn't work through cmd I tried running the install.bat itself and it still gave me the same error, weird
>>16983 yeah that is odd. hmm let me take a look to see what might be the issue
>>16983 >>16984 actually, when you ran the installer did you get an admin request for powershell? it's required to change the execution policy so that sd-scripts runs properly
>>16982 can I still use lora_locon_extract? is it integrated into UI? Thanks for the update!
>>16985 not really but I'll try running admin for powershell to see if it helps
>>16986 unfortunately no, those will be added within the next week or so, I was more focused on getting out the UI for the main section, training lora. however the old scripts do still exist, and I updated them to the newest version of sd-scripts. you can grab them here: https://github.com/derrian-distro/LoRA_Easy_Training_Scripts/tree/old-scripts just follow the install guide there if you don't already have the old version installed. >>16987 it should ask you for admin, and it does so for me without an issue, it's a bit hard to fix it when I can't replicate it...
>>16987 nah still didn't work even though I changed the Get-ExecutionPolicy with admin powershell and I still get the same error, I'll just stick with the popup scripts for now
>>16989 I don't know how it worked for you when you first installed the popup scripts, because it installed the exact same way...
>>16976 What's the difference between this and kohya
>>16991 you talking about sd-scripts or kohya_ss? because if sd-scripts, then it has a UI and makes training actually easy, if kohya_ss, makes the UI actually fucking usable, and removes gradio from existence, because I fucking hate it. so long story short, nothing beyond not being a shitty gradio UI and actually having proper thought put into it.
>>16990 weird I installed a separate folder for the latest version of popup scripts not too long ago. I'll try it out one more time to see if it still works
>>16993 weird it seems like that one didn't have the install.bat file, I just ran the installer.py file to install everything the last time I did it.
>>16994 yeah, I changed about 4 lines of code, removed the original git clone, added the pip installs for the UI and lycoris, then made an install.bat that is in the main folder which literally just calls installer.py. so it almost 100% is the same, and I didn't touch the execution policy bat file or the code calling it.
>>16994 k fixed it, seems like I didn't have system32's Powershell path to my environmental variables
>>16996 I thought that was in path by default... what?
>>16996 >>16997 either way, glad it's working for you now, it's 3am and I need to sleep, I'll be looking at the thread tomorrow so if you encounter any bugs you can reply here or open up an issue on the github (which I'd prefer that vs responding here)
Wasn't the case when I looked at the PATH for my environmental variables, weird shit but otherwise it's all good now. I'll probably post something if there's any issues regarding baking in the future but I had an error when trying to install 2.1.0 Torch >ERROR: Could not find a version that satisfies the requirement torch==2.1.0.dev20230320+cu118 wondering if that's because the latest devbuilds are running cu121, otherwise I don't mind because the latest devbuilds for Torch was giving me errors with outputting stuff for the SD-Webui
>>16976 Oh my god JC, a VIRUS
Someone has the Ishikei Lora? I have no idea why they took it down from Civitai
>>16942 >>16961 Is 128 dim bad now? I haven't kept up with this shit for like two months (stopped around the time people were experimenting with new optimizers and locon/loha/whatever) and I remember the meta being 128/128 or 128/64 for dim/alpha back then. Any new meta starting parameters? Is 1e-4 unet and 1e-5 tenc still a good starting point? What about the number of total steps per image in dataset (supposedly you need more steps the more images you have) with relation to batch count etc.?
>>17002 not really but the only time I had to use 128 was for a character that wasn't even human, I just then resized it to 64 with the resize LORA.bat as for learning rates pretty much yeah, nothing much has changed since the last time you baked something as for total steps most people are still aiming for 1400-1600 with 2 batch
When easy anon is up I'm confused, did the whole structure for training image folders change for the GUI? it doesn't instantly discover my img_dir folder.
>>17004 Yeah, now you pass it in each folder with the images in it, this is to facilitate the ability to change a bunch of things for it. Namely this change allows you to select a keep tokens per folder it also means you no longer need to name your folders <number>_<name> and can just name it whatever you want. >>17000 Oof, guess thats one less person that's gonna use my scripts >>17002 i have a personal hatred for dim128, i find that 99% of the time i get better results at mich lower dims, mainly because in a lot of cases dim128 breaks things like eyes, hands, and backgrounds. also in terms of learning rate, i actually use 5e-4 unet, 1e-4 te mkst of the tine, but i also dont bake one attempt, i bake multiple, starting at 1e-3 unet then change the larams based in output.
>>16960 Yeah, I'm saying thay higher dims learn more artifacts. This is because you reach a point where the weights don't need to be any higher, and it still has all this extra space thay goes unused, so it starts to fill that space with unneeded shit, noise basically. I've seen, multiple times, that lora look better, with less random artifacts or noise, at lower dims, it's honestly night and day. >>16961 I do that because at that point I know I can just make my own at dim32 or lower, and it'll probably be better. I've always called high dim the "lazy man's" way of training. It works, and you don't have to tinker with it, but the results are almost always subpar compared to effort being put in at dim32 or below.
>>17005 >>17006 God typing on phone sucks
>>17006 but learning at higher dim and then resizing is the way right?
>>17008 Never once saw that as the best option, no.
>>17009 it's incredible how contradicting opinions are on lora training shit
>>17010 I've pretty much always pushed for low dims because I've seen better results on lower dims. When it comes to training at a higher dim then resizing down, if thay higher dim is 32 and you are expecting to resize it down using a fro resize, then sure that's fine to do, but I haven't seen a good reason to even do that, I'll just train at lower dims and bake it another time if I think it didn't learn enough.
>>17011 That being said, I don't even train lora that much anymore, I mainly train locon. Characters get trained in dim16 lora usually, but styles are trained as locon, dim16 conv dim8 (Though I have a feeling that dim4 conv dim 16 will work better), because they learn style significantly better
>>17012 And i pretty much have only been training styles lately, minus Arona + plana
>>17005 I see. I've noticed some weird stuff like doubled details on outfits or bad hands. Maybe I could try retraining at lower dims a see. Also how do you decide if you need to change unet or tenc? I kind of understand what unet does, but never figured tenc out. I tried making x/y grids with different unet and tenc weights and tenc barely did anything, even at 0 the lora worked somewhat fine. Like it changed something, but what exactly?
>>16976 PS redacted\LoRA_Easy_Training_Scripts> python --version Python 3.10.5 PS redacted\LoRA_Easy_Training_Scripts> .\install.bat Running on windows... ERROR: You don't have python 3.10 installed, please install python 3.10, preferably 3.10.6, and add it to path Uhh...
>>17014 it's pretty intuition based, I've trained a lot of test models during my research into lower dims, so I just have an idea of what to change, give me a bit and i'll try and give you a general write-up of what I do to train. >>17015 oh you know what, I think I have the check not allow it if the python version is less than 3.10.6, I believe that was a mistake, let me fix that up
(423.11 KB 1061x1500 103301503_p0.jpg)

Anyone know if there's a lora already of this artist (fushichiyo/cockchiyo)? I couldn't find one but I don't know if they go by other names so I want to make sure before I commit time to it since they have some heavy censoring. https://www.pixiv.net/en/users/7575/artworks
>>17011 >>17012 >>17013 What are your usual training settings? I've seen it mentioned on cuckchan that training at 128dim and resizing with some tool is superior to simply training at lower dim because resizing "trims the fat" while low dim training simply underfits
>>17018 different for every model. I bake multiple times to figure out the best dim for it, and pretty much never do I need to go higher than 32, and even then that's not even true most of the time. besides, you can see just how ludicrously too big dim128 is if you resize at all, like even a fro at 0.99 max dim 128 cuts the file size in half usually. which means it completely cut all of the weights in half. I just think it's because people are lazy and don't want to put effort into their models.
>>17019 well, I bake multiple times to figure out the best everything for it.
>>17017 That's Feral Lemma. Someone on this board made LoRAs for him a while back.
>>17021 Wow I'm retarded, I thought it was somebody else with a similar style because of the name difference but I just checked the tag wiki. The feral lemma lora is pretty good so I guess I don't have to deal with that.
(2.32 MB 1024x1536 ComfyUI_05692_.png)

(2.35 MB 1024x1536 ComfyUI_05701_.png)

>>17015 updated the scripts, fixed that bug + more. changelog: made the training threaded, so that it can be done while still allowing you to use the UI fixed the way betas was being saved in the toml files so that it doesn't break on load fixed the installer by removing torch 2.1.0 and adding 2.0.1 in it's place as well as add a fallback execution policy bat that just assumes your powershell is at the normal location in system32, also fixed the version check to only check that you are using some version of 3.10 Fixed the limit on the samples input box to allow higher than 99
>>17018 they're different things. Resizing is kinda a way to save a lora/remove some overfit. See pic rel on how it should be set - had similar experiences to the author with 128 being too much but 64 too little, so the 96 he mentions is probably the peak. The 32 dim and below schizophrenia was laughed out of the chan proper months ago along with the guy doing flawed testing and drawing wrong conclusions.
>>17025 I don't understand how people manage to be unable to make character lora that work well at dim16, hell I've managed to do it more than once, and I never have issues with reproducing the character. the schizo is dim128. And I will never change my mind on it.
>>17026 You're missing the point. Low dimension is incredibly easy to bake, its just that it ends up "underbaked" in an unique way - its missing data because of said dimension. Higher dim loras are naturally better but also harder to make because you can actually overfit them. Basically, uhh, skill issue dimschizo. Denying reality is not very healthy when we had grids showcasing low dim effects on loras since early days.
>>17027 the fuck are you on? the exact opposite is true. dim16 lora aren't under baked, when you know what the fuck you're doing.
>>17028 They are and its impossible for them not to be, because they're limited by the dimension size no matter how high you set the LR
>>17029 but how am I able to get entirely not underfit lora at dim 16 if that's the case?
Unrelated to anything, I love the cuck tutorial on CIvitai that has people tagging their training data like "a woman wearing a STYLE dress" or whatever so it literally doesn't apply jack shit unless your prompt is written exactly like their schizo ass
>>17032 wouldn't that help if you also combined it with wd tags? like train with both blip and wd tags?
>>17033 I mean, sure, but I'm talking like, every image has exactly -one- tag and it's the entire description of the image
>>17030 You simply dont realize its underfit because you settled and never compared.
>>17034 oh I see... yeah that's 100% useless
>>17035 but it replicates the character perfectly, what's the point in going higher when it works perfectly at lower?
>>17037 Try it and find out? Its easy to say that something is "perfect" if you never compared it to anything else, some issues or missing things only become visible that way. There's a reason both guides and people recommend such high dimensions and its not just "everyone except for me is wrong" lol.
>>17038 whatever, i'm not fighting you on this, you're wrong and so long as your fine with being wrong then whatever
friend is looking for a foo fighters (the jojo stand) lora, anyone got one? and for a decent jolyne one
not sure if it's possible but would it be possible to add a form of queue training to the Easyscripts GUI? not sure if just putting a bunch of TOML files into a queue button would help with that
>>17042 also looking at the toml file for my config through notepad I notice the min_snr_gamma option is no longer there also on the GUI, did it turn out to make LORAs worse or something?
>>17042 already planned, just had to get rid of it for the time being because I need to figure out how I want to set it up. >>17043 ah shit, of course I forgot something
>>17023 Pretty nice, catbox?
Is setting max tokens to 225 better than 150 for LORAs?
>>17046 it only affects the max length of a caption file, unless you have stupidly long captions, it doesn't really matter
>>17043 Alright, I added the option
>>17047 ah so not so much of a major effect on the end product, alright thanks for the info
>>17041 I don't even understand how to inpaint. It seems hard to do on collab since you need decent RAM size.
>>17023 What a man of culture, catbox please?
>>17041 When you use inpaint do you delete all your prompt and only write what you want changed on the masked part? I'm new to this
>>17052 generally yeah
>>17047 So why is it not set to higher by default?
Been out of the loop for a long while. It seems colab killed A111 for free users? Started messing around and apparently ComfyUI still works like a charm so I'm forced to learn this new UI if I want to continue to generate stuff without buying a GPU or a whole new PC.
>>17054 I suppose it's due to the fact Kohya had his default max tokens set to 150 for sd-scripts
>>17055 I would also like to add that whoever named it ComfyUI has a sick sense of humor because this isn't comfy at all.
I'm getting 8dim pilled, some of my character LORAs are looking better than the higher dim ones and get the details as long as if I play with the exponents correctly, I'd probably get shit on people who stick with the basic 128 but my character LORA has her details so what the hell is even the right dim
>>17058 The correct answer is "it depends" and you all are creating idiotic discourse over nothing.
>>17055 >ComfyUI still works like a charm It does? I thought they nuked any sort of remote UIs. I'll try it then but yeah it does seem fucking autistic.
>>17060 Yup to both counts. It does work and it is hilariously fucking autistic. Have fun relearning everything!
>>17061 To add to this: Shit's so autistic that I doubt this would catch on with the average joe so it shouldn't be a burden on google's servers. And ComfyUI as a whole is pretty unpopular so we should have our free solution as long as it stays so.
>>16975 catbox? loving the detail
>>17052 I usually leave the prompt as is if I'm inpainting with the inpaint area mode set to whole picture, except when it adds stuff I don't want, then I remove that part from the prompt or add a negative if it's something not part of the prompt. If I inpaint with masked only (inpaint at custom higher res), then I strip everything from the prompt that's not relevant to the masked area + padding.
Any good new models lately? I haven't prompted in like 2 months and am still on old stuff like Based64 and AOM3
I notice the tagging on some Civitai preview images are literally the LORA and the keep token,I assume that's due to pruning appearance tags?
Will chaning the name of a Lora break anything when used as a prompt? I'm trying to organize my Loras and I hate the ones with gibberish as names
>>17067 can you share your prompt to give it a try? Loving that belly
>>17068 no it doesn't the only way a LORA is affected once it's finished baking is by using resize scripts or block editing
>>17065 I'm gonna shill >>16730 again, because I like it.
(2.03 MB 1024x1536 cum.png)

>>17069 Catbox for the "base gen": https://files.catbox.moe/86e3ct.png What I used for inpainting: nsfw, scan, ((traditional media)), (sketch:1.2), (extremely detailed CG unity 8k wallpaper:1.2), soft lighting, HDR, (best shadow), BREAK penis, cum, overflow<lora:artist-zankuro-new:0.3><lora:ponsuke_(pon00000):0.8> <lora:pregnant_dim64:0.9> Negative prompt: (worst quality, low quality:1.4), (greyscale, monochrome:1.0), The image posted is the image I inpainted so yes I cropped some cum from a different gen but you can probably just draw some cum in paint or something.
>>17073 break penis cum
>>17074 Yes, the day of the average prompter.
>>17073 >Ponsuke Lora Where do you find those loli artist Loras? I don't think they'll be on CivitAi
>>17076 The Ponsuke Lora can be found on anonfiles, the link can be found in one of the previous threads. I'd dig it up but i'm supposed to be working.
>>17076 gayshit
what were the main signs of overfitting? the keep token details of your LORA showing up when you didn't use it in your prompt?
What do I need to change to reduce style being learned? I'm trying to train a model on a small dataset but the style seems to be prevalent despite anything I change. I've tried lowering the learning rates, lowering the netdims, and lowering the training steps.
>>17080 the dataset >>17079 kinda?
>>17081 fug, I'm not sure I can do anything about that then. I could only find like 8 images of the character I'm training.
Man, I was excited for the 4060 Ti 16gb potentially being a good card for Stable Diffusion, but it just looks like it's going to be absolute trash value. How's SD on AMD these days? Is it even possible to run it on Intel Arc?
>>17082 You could try using various models and loras to make images in different styles and then add the best ones to your dataset and train again.
>>17082 >>17084 Or maybe even try using controlnet (I think it was the shuffle one?) to make images in different styles for your dataset.
>>17082 >>17085 Shuffle could work! I think there's also this new controlnet option called "reference", could work pretty good as well
>>17081 I suppose less steps can fix that >>17083 was the memory bus' size actually bad for AI?
>>17087 Rasterization performance is withing a couple percentage points of the 3060ti. At 1440p + the 3060ti even outperforms it in some games. The memory bus' size causes some problems when loading textures too, but idk how much of a problem that would be. Dunno for AI, but in general it seems to be a horrible value of a card. I know my excitement for the card has been dampened a lot
>>17083 >SD on AMD lol hahahahahahahhaahahahahahaha No. >>17087 Maybe. Didnt you save epochs to check?
>>17089 the thing is epochs are weird to the point that sometimes the semi-final epoch looks crap compared to the previous one, not sure if that's randomized nature or some shit
>>17090 Its called cosine with restarts performing a restart. And I dont see how this matters, if you saved epochs you'd be able to test them and see if it's overbaking or something else
how are there still newfags itt? Did you guys go recruiting elsewhere or something? Considering this was originally refuge of semi oldfags its weird to see people still asking elementary questions here
>>17092 >how are there still newfags itt? It's been on the gitgud repo for the last 3 months or so.....
>>17071 Tried this for a bit, it seems worse than the ComfyUI collab. >Pushes all SD files to temporary colab storage meaning you need to redowload LORAs, models, embeds and everything every single time you want to do something which is the main dealbreaker >Clusterfuck of prebuilt extensions. Starting it for the first time was like loading up one of those old browsers with 10 installed search bar extensions >Literally everything is pre-built for you instead of giving you a clean SD install then you install the things you want/need like with the LastBen collab Upside is that it is using A111, meaning that you can use popular extensions and the things everyone is using unlike with comfy where you're at the mercy of the ComfyAnon updating his colab/UI whenever something happens. Downside is that it's notably worse and having to go through the entire setup process of setting up models, loras, vaes and embeds every session is a major deal breaker.
>>17094 What does comfyui currently lack compared to A1111 with all extensions? Is it still good if you are able to get through the whole autism?
>>17083 its NOT worth it that thing was made for CUDA and you while you can use AMD for Stable Diffusion itws not as good as Nidia t.AMD user
>>17095 Honestly? I don't know. I only started messing around with it. My first impression is that it seems better in some ways since you can see how the image is actually being generated and fine-tune your image better than A111 since you can customize your workflow and see what's happening in the background in a better way but it's also way harder/autistic to get into it if you're a casual user like me over something like A111 that's pretty easy to get the hang of. Your options are basically to either get used to ComfyUI if you're a colab user, deal with the annoyance of having to set up your colab every single session for the privilege of continuing to use A111 or find someone who can modify either collab to work without the google warning of "Your VM's life ends 30 minutes from now".
Nevermind i'm beyond retarded and sleep deprived so I have deleted my stupidity. Just run this on the nocrypt colab from distutils.dir_util import copy_tree copy_tree("/content/gdrive/MyDrive/sd/nono-folder/models/", "/content/sdw/models/") copy_tree("/content/gdrive/MyDrive/sd/nono-folder/embeddings", "/content/sdw/embeddings/") nono-folder is my old SD installation on that particular drive and where all my models, VAEs, LORAs and shit are and it's copying them to colab's local storage. Depending on how big your folders are, this will probably take a while but it automates the entire process and skips having to learn how ComfyUI works.
did that anon who was making the ufotable finetune ever release it? i remember reading about it months ago and wondered how it turned out
>>17101 >One of the anons from here actually made the ComfyUI infinitely better I kneel
>>17101 Do you have any tips on genning on ComfyUI vs A111?
>>17103 a couple I know of - ComfyUI's algorithm for prompt keyword weights is different than A1111's. in practice i find that you don't have to go past 1.3 for something you want to emphasize. more details here: https://github.com/comfyanonymous/ComfyUI/discussions/473#discussioncomment-5604052 - if you want to use A1111's weighting algorithm anyway, have a look at https://github.com/BlenderNeko/ComfyUI_ADV_CLIP_emb - reroute nodes are indispensable for cleaning up wires. they act as an inflection point in the middle of the wire you can drag away. - ControlNet nodes are included in base ComfyUI but you'll need to bring your own preprocessors if you want to use random images with canny/hed/etc. https://github.com/Fannovel16/comfy_controlnet_preprocessors - the tiled ksampler node lets you denoise huge images by splitting the latent into tiles and denoising them separately. https://github.com/BlenderNeko/ComfyUI_TiledKSampler - have a look at the official examples page for some common workflows, they go into more detail than i could. https://comfyanonymous.github.io/ComfyUI_examples/ - more advanced workflows can be found here: https://github.com/wyrde/wyrde-comfyui-workflows - right click an image output on a node and click "Copy (Clipspace)". click the "Clipspace" button in the sidebar and you can edit the copied image's mask. then right click a "Load Image" node and select "Paste (Clipspace)" to use the image with the mask. now check out the example inpainting workflow for how to use this. - their matrix chat linked in the readme is pretty useful since you can chat with comfyanon directly and ask them questions with a quick turnaround. - you can turn any widget on a node that displays a value into a wireable input by right-clicking the node and selecting "convert <X> to input". then add a `utils -> Primitive` node and hook it up to the new input. - if you're disheartened by nodes you can try ComfyBox (its default workflow is an A1111 clone), but be aware as with ComfyUI that it's still alpha quality software and your saved workflows might break.
>>17104 Does Comfy have booru tag autocomplete?
>>17105 not currently, but it's something I'd add to ComfyBox at some point
>>17076 I installed torch2 and now I can't gen. I fell for it
>>17101 this looks like the only realistic alternative to auto1111 i'd consider. thanks based catbox script anon
>>17107 if you didn't know what you were doing why didn't you make a back up?
>>17106 Ah, that's kind of a deal breaker for me. I like what it can do and the node system looks intuitive (maybe because I have some experience with Blender nodes), but booru tag autocomplete is too convenient for me to drop as I use it to search what tags are even available when I type a word and how prominent that tag is, plus links to wiki in case I'm not sure what it does. One thing I want to try in Comfy eventually is setting up double hires fix, where the first one uses latent at some smaller scale to add details and then a second normal one.
>>17110 >One thing I want to try in Comfy eventually is setting up double hires fix, where the first one uses latent at some smaller scale to add details and then a second normal one. NTA but isn't that already in? I'm pretty sure I saw something like that in the GAQ
>>17105 >>17110 damn I haven't ever used autocomplete and manually check boorus for the tags...
>>17111 Well yeah, I meant that it's something that sounds useful and that A1111 can't do automatically in one go. >>17112 Honestly, the default setup of the addon is kinda shit, so I set it to show 30 tags and changed the css so the spacing isn't so huge and I don't have to scroll to see the full list.
>>17112 >>17113 I just have the tags I like memorized and find autocomplete to be a bother personally
>>17114 Yeah, sometimes it's just faster to type them out manually, but it's useful for all the different "cum on X" tags and similar, including aliases, so for example "cum on face" automatically turns to "facial". And as I said, it's also good to check if a tag is even worth using by seeing the total image counts. Also helps with quickly adding embeddings, especially all the different negative ones, as I never remember if they use - or _ in the filename.
>>17109 I don't appreciate your tone
it's a valid question
>>17117 meant for >>17116
can anyone explained why epoch 10 of a 14 epoch lora looks nothing like training the same lora to stop at the 10th epoch? i understand shuffle captions means they're never going to look identical, but the difference is staggering
>>17119 Scheduler probably. If you stretch a curve from 0-10 to 0-14, the Y value will simply be different for X =10.
(1.09 MB 1024x1024 00178-4011820859.png)

(1.11 MB 1024x1024 00194-2813807634.png)

(1.09 MB 1024x1024 00176-262294490.png)

>>17121 i don't believe kuroko would be mad in this situation
>>17119 restarts happen at different points as >>17120 said. Are you capable of thought bro? Cause that's pretty straightforward.
>>17120 ah yeah thanks that makes sense >>17123 kys
>>17123 oh yeah i forgot >--lr_scheduler=cosine so actually kys
>>17125 >>17124 >>17123 just use constant
>>17126 no one asked for your input
huh?
>>17129 i misread your post. either way, guy who developed comfybox is catbox anon, so i appreciate him
>>17130 Is he? I always thought they were two separate anons
>>17131 catbox archive anon* though i wouldn't be surprised if they were the same one
(1.29 MB 960x1152 catbox_7r4jdv.png)

(1.05 MB 960x1152 catbox_w7vwaf.png)

(1.27 MB 1152x1152 catbox_32cl5g.png)

(1.04 MB 1152x1152 catbox_imzekm.png)

new style locon: kuroshiro/@kuro96siro46 https://mega.nz/folder/OoYWzR6L#psN69wnC2ljJ9OQS2FDHoQ/folder/rtw3zIaI non-h artist but has a nice blocky style that may mix well with others
>>17132 yeah i'm not catbox anon for the record. added danbooru tag autocomplete but it's pretty basic. to test it out download the latest version, unlock the graph, click "toggle props", select the prompt textbox and in the "variant" dropdown choose "code"
>>17131 We are indeed separate but it would not surprise me if comfybox anon has seen my work in other circles
>>17135 actually I did use your PNG metadata scraping code because it's much superior to ComfyUI's buggy version. if you don't mind that is.
>>17133 Thanks. How many restarts do you use for training?
(1.61 MB 960x1344 catbox_57vhk6.png)

>>17138 I use cosine with 4 restarts with an epoch count divisible by 4. Seems to help quite a bit with preventing overfitting. You can check the metadata for the rest of my hyperparameters.
>>17139 >I use cosine with 4 restarts with an epoch count divisible by 4. Seems to help quite a bit with preventing overfitting. Thank you. >You can check the metadata for the rest of my hyperparameters. Yeah I checked those, but it doesn't specify amount of restarts in the meta
>>17137 I had to double check your repo to see what you meant, yeah that's whatever with me. I took almost all of it from this anyway, which borrows from yet another source. I think even the webui's code is a bit messed up for it and I haven't bothered looking into fixing it (doesn't always seem to filter certain useless keys like "dpi" when it should be) https://github.com/moonshinegloss/stable-diffusion-discord-prompts https://github.com/hughsk/png-chunks-extract
>>16977 catbox?
>normalfaggots seething over AI gfs I need open source text to voice AI now
(1.25 MB 1024x1024 00013-2940187975.png)

>>17122 she's mad because you walked in
>>17144 wait for real? normalfags are barely finding out? kek
what is charm22 and where tf did i get this lora
>>17101 >>17134 Is it possible to use ComfyBox using colab? Here's the official colab by the dev for reference https://colab.research.google.com/github/comfyanonymous/ComfyUI/blob/master/notebooks/comfyui_colab.ipynb
>>17148 i have not tried since i can't be arsed to fuck around with colab and my storage is almost full, but it should work in theory. i might try looking into their official colab to see if i can adapt it.
>>17146 it's mostly the /vt/ shitpost >PARASOCIAL FUCKS and all that jazz, normtards just really hate lonely guys that find alternative forms of affection to be honest.
>>17150 If you gave a single penny to a roastie that pretends to be anime girl (vtuber) you are fucking retarded. But vtuber anime girl designs are cute yeah
>>17150 Oh its shitposting ok there's a good crossover audience between AI gen community and the vtuber one so it makes sense. >>17151 Listen man, im just gonna fap to their avatars getting railed by a bunch of fat ugly bastards, I ain't giving shit to no roastie,anime avatar or not.if poeple wanna "donate" money to them then whatever.
>>17152 i'll jerk off to mori getting railed by a horse but i'm NEVER gonna watch her streams, eww
>>17152 >Listen man, im just gonna fap to their avatars getting railed by a bunch of fat ugly bastards, I want to fap too but holy shit it is so hard to generate good sex using AI
>>17148 >>17149 took longer than I'd hoped but here it is. the setup process is more arduous than ComfyUI's notebook but i've at least verified it works. https://colab.research.google.com/github/space-nuko/ComfyBox/blob/13de9b56093ce95831b0f25b6d7b8eaf3cb85e42/notebooks/ComfyBox_Colab.ipynb if you get fetch errors even after setting the backend URL then be sure to clear your cache (the config is fetched over the network so it might not get updated if you just refresh the page)
(1.78 MB 1024x1536 00111-670157322.png)

(1.75 MB 1024x1536 00017-3419354432.png)

(1.80 MB 1024x1536 00020-3340629476.png)

(2.11 MB 1024x1536 00003-1153529287.png)

https://mega.nz/folder/ctl1FYoK#BYlpywutnH4psbEcxsbLLg redid hades style lora. much better than the 994it one imo and fixes some problems with the old 1134it version. moved the older versions to archived folder
(1.66 MB 1024x1536 00010-2036489106.png)

(1.95 MB 1024x1536 00223-1056919676.png)

(2.02 MB 1024x1536 00141-1178311093.png)

(1.84 MB 1024x1536 00028-4083798155.png)

did any new cool updates roll around or are we still dealing with the same shit?
any tips on how to deal with the third nipple when prompting pregnancy?
>>17158 Controlnet 1.1. Some more loras. Nothing revolutionary, but what do you expect? There's only so much you can do with the shitty base underlying model
(3.32 MB 1280x1920 catbox_8mmhdp.png)

(2.82 MB 1280x1920 catbox_z82fqh.png)

(3.08 MB 1280x1920 catbox_94q95s.png)

(2.73 MB 1280x1600 catbox_2coliy.png)

(1.54 MB 960x1296 catbox_8532do.png)

(1.31 MB 960x1296 catbox_cdj6bp.png)

(1.54 MB 960x1296 catbox_8sup52.png)

(1.86 MB 960x1296 catbox_8kylp5.png)

>>17162 kinda wonky but I can get some pretty cute results, thanks anon
>>17158 Same Shit Different Day we'll be forever confined by the shackles that are the limits of NAI
(882.10 KB 1024x1536 00052-3832248076.png)

Does this look blurry?
easyanon any reason why your scripts are giving me errors like >TypeError: ClusterConfig.init() got an unexpected keyword argument 'downcase_fp16' when I try running >accelerate launch --gpu_ids="1" main.py --popup with your scripts now?
>>17166 you are using the older scripts? not sure.
(3.12 MB 1280x1920 catbox_k7rk7p.png)

(3.12 MB 1280x1920 catbox_5gh3fy.png)

(4.16 MB 1280x1920 catbox_ga595s.png)

(4.16 MB 1280x1920 catbox_wci4v4.png)

>>17162 Well I took the time to de-jpeg and resize as well as prune and add some more better quality images and the result is almost exactly the same. Even left out the character name tag and removed keep_tokens. Had to double check the metadata just to be sure
>>17167 nah I also got it with the GUI
(2.86 MB 1280x1920 catbox_tepssa.png)

(2.89 MB 1280x1920 catbox_kb76vr.png)

(2.95 MB 1280x1920 catbox_qyyy5a.png)

>>17168 And then slightly different training settings can get you wildly different results
>>17169 it seems like only a nippon dude has dealt this but through the webui, weird >solved by removing downcase_fp16: 'NO' line in C:\Users\<username>/.cache\huggingface\accelerate\default_config.yaml
any good hentai monster pack models besides the barebones stuff on civit?
>>17171 oh that is odd, I guess it might be a GPU issue? hmm...
>>17165 catbox please?
Any fix if I keep getting constant black images even with no half vae?
Has anyone made any lora for covered in cum or bukakke? Because I swear the AI always absolutely refuses to draw any cum on the girls body
>>17177 there's a few on Civitai but I don't really trust their shit to be quality anymore.
(1.16 MB 1024x1536 00006-3600817910.png)

(1.19 MB 1024x1536 00010-3183483921.png)

(1.72 MB 1024x1536 00019-1276705870.png)

(1.67 MB 1024x1536 00172-909234093.png)

(1.69 MB 1024x1536 00020-1102049231.png)

(1.44 MB 1024x1536 00289-1064776240.png)

(1.67 MB 1024x1536 00172-909234093.png)

(2.06 MB 1024x1536 00500-3104849345.png)

resident noob here: do you hardcore guys go deep into inpainting, or is it all about the txt2img for you? the amount of inpainting and settings i have to tweak is kinda getting annoying, so i just wonder what an actually good workflow looks like. i typically just let txt2img gen all night (slow pc), prune the results the next day, and spend for fucking ever trying different masks, mask blurs, prompts, denoises, etc all while it's ultimately an inpaint for something minor (but still important)
I've been revisiting old loras and stuff for fun I wonder if the anon that made this still browses this place or /hdg/
>>17181 Well my current workflow is text2img>variation seed rolling/GIMP>img2img>inpainting. I only started taking inpainting seriously recently.
(1.70 MB 3000x3200 Lama.webp)

>>17181 >>17184 I only inpaint if I really care about a pic by copy pasting the bitmap into photoshop, edit it there using the eyedropper tool to pick nearby colors and a very small brush (1-3px) to paint over stuff, along with other tools like content aware fill to outright remove stuff, clone stamp tool for textured surfaces, smudge tool with scattering to blur seams or to make smooth transitions between colors etc. Then I select the whole canvas and copy paste that back to inpaint, draw a mask and let it regen my painted over bits, one at a time. I repeat this process multiple times, each time with smaller denoise, until I'm satisfied with the final result. Also as I paste from webui into photoshop, it always makes a new layer, so sometimes I erase a part of the new layer to blend it with a previous layer if I don't like a part that I thought was better before. The reason for using a small brush is that I find it blends better with colors instead of making these huge flat color areas. Anyway, I recently found this jap guide about using lama cleaner to quickly remove unwanted stuff before you inpaint. Might be interesting. https://github.com/Zuntan03/SdWebUiTutorial/blob/main/_/doc/LamaCleaner/LamaCleaner.md
(1.67 MB 1024x1536 catbox_6okpak.png)

(1.87 MB 1024x1536 catbox_d7tri3.png)

(1.67 MB 1024x1536 catbox_0j1dnm.png)

(1.72 MB 1024x1536 catbox_x9jpal.png)

Any Gobbo enjoyers here? Currently looking for some inspiration for my favourite shortstacks
(1.35 MB 960x1152 catbox_rjy0vp.png)

(1.37 MB 960x1152 catbox_dnx44c.png)

(1.37 MB 960x1152 catbox_vu7xms.png)

(1.36 MB 960x1152 catbox_6q900l.png)

>>17186 I really like this noise (artist) LoRA, I got it off civitai but it looks like it's been taken down. Does red and blue onis really well too. https://files.catbox.moe/t2szzp.7z
(1.22 MB 1024x1536 00051-1129270376.png)

(1.33 MB 1024x1536 00054-443063568.png)

>>17188 >>17189 Cute bellies
(1.71 MB 1024x1536 catbox_fg7p47.png)

>>17187 looks interesting Gonna test it a bit
>>17184 oh i looked at the variation seed roller thing briefly and figured it was a cheap gimmick. i'll give it a try >>17185 sounds like a lot of the manual tedium of what I'm doing currently tbh. does lama cleaner work better than SD at removing objects or something?
>>17184 what's variation seed
(16.79 KB 912x184 seed.png)

>>17192 I basically use a variation strength of 0.01 when I'm trying to fix minor stuff but do play around with it. >>17193 A way to "reroll" a gen, the higher the variation strength is the bigger the changes will be.
>>17192 >does lama cleaner work better than SD at removing objects or something? No idea, but from the looks of it, it looks like a better content aware fill, and it's faster.
>>17185 i forgot to ask: do you have any parts of your workflow that are automated? for instance, i've currently been working on folders of wildcards for things like the thick curvy archetype or the loli archetype. each of those folders will in theory have various prompts to achieve said archetype, and i'll let them run while i'm at work so i can compare what works best for a given seed/character/whatever i also want to begin creating wildcards that contain different CFG + denoise + step counts to run my chosen seed through to see what the sweet spot is for any given gen without me needing to manually do a bunch of that by hand. i know that xyz prompt has something like this, but it doesn't allowing saving this stuff in any direct manner, and it has limited flexibility
>>17196 i keep adding stuff on but an additional followup questions since you guys are the most likely to know: is there any way to cycle through different masks since how you mask something can have such a big impact on how the inpaint is done? i'd much rather have a folder of masks to try overnight for a given inpaint rather than having to input a new mask each time by hand
>>17177 >>17178 You could probably train a better LORA yourself tbh considering theres a fuck ton of material with that subject out there
anyone upscale small images? I'm trying to look for something that can get rid of jpeg artifacts or noise after I upscale with the extra tab in the webui
>>17199 I found a few denoising models on the upscale wiki, I'll see if they help
>>17071 Anyone else using it? img2img seems to be completely fucked, is there any extension that could cause it?
hey all, did some bug fixes and added queues to the UI. the queues aren't as polished as I wanted it to be, but it's still pretty nice looking. I believe they work, and I tested them enough that I believe it will work, but you know how technology is. anyways you can update via update.bat if you have it, or git pull then running update.bat https://github.com/derrian-distro/LoRA_Easy_Training_Scripts
>>17196 Nope, I do everything manually, I'm not really that hardcore and I only gen when I have a specific idea or when I want a quick fap. I write my prompt and gen one at a time, each time I decide if I save the pic in my favorites or not and gen another. After a few gens, I modify the prompt slightly to either fix/improve stuff or to get something new. If I really like a pic and want to share it with people, but it still has defects, I try to fix it with inpainting, either right away or sometime in the future. >>17197 I know you can load a separate black and white image as a mask, but you'd still have to manually replace them. ComfyUI would probably be better for automating stuff like this, I think it allows you to queue different workflows you set up in advance. But I've only read the guides and readmes and haven't used it yet.
(1.32 MB 496x496 uoh.gif)

>microsoft adding AI to windows yeah I'm not letting Bill look at my shit through his AI, time to switch to linux
might be a useful voice tts AI, saw it on /g/ I'll try it out, 10-30 second voice clips https://github.com/gitmylo/audio-webui
so d-adaption v3 exist https://rentry.co/dadaptation-for-lazy-people >With the release of V3 and other versions of the DAdaptation optimizer, I firmly believe that this is the best optimizer to use and oust AdamW8Bit as the default optimizer
>>17208 unless Bark has somehow magically improved the results are going to be terrible
>>17210 the results I heard are still a bit robotic, we'll never reach closed source voice AI levels for awhile
>>17211 Robotic isn't the issue with Bark when I used it, it's that it hallucinates stuff or doesn't even follow the input text in the slightest. Might as well use RVC and speak the lines myself.
>>17207 boy I love it when cropos rear their head in on something they didn't need to, cant wait for shit like "SEE WHAT ILOVESHITTINGDICKNIPPLES69 IS CURRENTLY GENNING LIVE ON HIS AI"
>>17207 >>17213 you could just... not update to w11? keep old isos in case you need to reinstall? turn off windows update? not updating is always the best choice no matter what cybersecurity nuts tell you, their fearmongering tactics are just that. t. win xp & 7 enjoyer in 2023
>>17214 >win xp & 7 enjoyer in 2023 I really-really don't understand why wouldn't you use loonix at this point. Just fucking why?
>>17214 does win 7 even support all the shit you need for a1111 also go to linux like the other guy said, ffs. the cybersecurity crowd are right. security is kinda their thing lol
>>17214 >not updating is always the best choice no matter what cybersecurity nuts tell you, their fearmongering tactics are just that. dumbest shit i've read today. look i doubt anyone here is stupid enough to visit questionable sites or fall into boomer traps, but support for windows 7 has ended and there will no longer be any security patches, which is 100% not fearmongering and just cold facts
I still wonder how that anon did that devolver-chan lora with 8 images without overfrying it, it's still the LORA that fascinates me till this day.
>>17218 lazylora anon here. it is technically fried, though not badly. the character is simple. i am still planning on making a verison 2 with the gens i've added, but i'll need to stop being lazy and finish tagging the dataset
>>17217 >visit questionable sites you underestimate the power of cumbrain mind state
Did twitter go full retard mode? One of my fave Lolicon artists got suspended and had to remake their account with extreme censorship on their works.
>>17215 >>17216 >>17217 linux is for trannies also picrel
>>17216 >>17215 >linux but muh vidja gaems
>>17223 kys faggot >>17225 loonix is surely more compatible with modern vidya than fucking xp and pretty sure it'd be better than 7. that tard is likely baiting though
>>17226 >and pretty sure it'd be better than 7 7 EoL supports dx12 cope, seethe & dilate
>>17227 Are you the polshitter?
>>17229 he never left
I keep trying to prompt more softcore things or cute things but I keep falling back to the onahole lora and shit so degenerate that I can't post it anywhere
>>17209 Despite its release, I still haven't managed to get d-adapt to work for me. Always either undercooked or baked to high hell, never anything else.
>>17231 Post it on /d/
>>17223 obvious troll since you'd be on a newer windows if you gave a shit about games/game performance. also dual booting is thing for when you have no choice but to venture into corpo spying land
>>17234 nta but if you're on win11 you're retarded. its that simple. XPchads at least have a decent reason to stick to it even if its not ideal
>>17225 >nta yta. what reason is there to stick to an OS with swiss cheese security
Wonder if the tagger extension is going to get any new models that exclude copyright tags like the early versions.
4chan is complaining about easyscripts again, I suppose it's due to the new .toml format? Eh I'm fine with it, I just have my own input to put in but I tried out the queuing system and it just kept doing the settings that were left on the last .toml training settings I opened and set to the last training settings to be queued despite the first queue I set being a different toml. I'm probably doing something wrong in that case, I'm also wondering if block weights can be implemented into the GUI that's all I have to input so far.
>>17238 yeah, I fixed that bug already, it's in dev. block weight will be added in the next day or two
The latest update for the webui seems to change a lot, token merging or some shit but I heard that there's been some changes that has been making LORAs look worse, great another folder for an update thanks Voldy. >>17239 thanks a lot anon.
>test DPM++ 2M SDE Karras >looks fried as fuck did normies hype up this shit?
>>17242 they did?
>>17243 It's the shit that's been most hyped up about the latest update outside of the token merging shit
>>17240 Can't you turn that off? The commit stated it's turned off by default except hires fix iirc.
>>17242 Is that the thing you had to manually add before or something else? Because I remember using the other DPM++ 2M Karras Test and that one was good, it had less issues with continuity and generated stuff in a more correct way, but made pics slightly desaturated.
>>17246 it's a different one that some people added manually before the latest commit but yeah it seems like it was something like the DPM++ test models and now Voldy has added it to the list of samplers (the test models still have to be added manually however)
>>17245 it's off by default but yes you have to set a ratio yourself for that in the settings however I heard the speed boost and the outputs looking worse is not worth it. It's better to see what happens when Voldy is able to add TensorRT to the webui for speed boosts
>It's better to see what happens when Voldy is able to add TensorRT to the webui for speed boosts the fuck when is this happening? if this is real, VoltaML is probably in shambles lol
>>17248 >>17250 >Loras are baked in into the converted model >This takes very long - from 15 minutes to an hour Does that mean you'll have to make a new model for every lora combination? That sounds kind off terrible.
>>17251 it does but maybe some anons have sorta settled on a certain lora combination and would just gen industrial amounts with it now that they can do it even faster than before.
(206.49 KB 269x415 chrome_2023-05-29_13-14-36.png)

Why is A1111 not applying my VAE when I'm inpainting?
>>17253 make sure the upscaler is set to lanczos, GAN upscalers can fuck with the colors. if it still looks bad you can try enabling color correction in settings.
(3.07 MB 1280x1920 catbox_c91l2o.png)

Petite senpai...erotic...
Jewidia's CEO seems intent on price gouging until morale improves so start buying those used 4090s the moment people with too much money are gonna sell them for pennies once the 5090s drop https://youtu.be/IlWT_TdOK6s
(13.76 KB 181x181 1683192786714170.png)

>>17256 UOOOOH
(1.60 MB 1000x1496 catbox_xiq298.png)


Trying controlnet and img2img but struggling with extra eyebrows
(3.21 MB 1280x1920 catbox_7tblgo.png)

(2.66 MB 1280x1920 catbox_seui48.png)

(2.59 MB 1280x1920 catbox_1kppy7.png)

That powerstance though.
>>17258 how cheaply does used prior gen sell on average during a new launch?
>>17253 run webui with the --vae-path /path/here argument to ensure it's actually being loaded before troubleshooting further
>>17255 It's inpainting, as far as I know it's not even using an upscaler That setting didn't seem to help, at any rate. Since I've only tried inpainting hiresfix'd images, I thought it might be that I was getting the correct colors IN the mask and it was the outside colors that were wrong, but I got the same thing on a non-upscaled image too, so.. I'm baffled. >>17263 Nah, the VAE is definitely working in txt2img. Something's up specifically in inpainting, and I cannot figure it out.
(1.77 MB 1024x1536 00051-2621574765.png)

(1.81 MB 1024x1536 00043-2456255249.png)

(1.97 MB 1024x1536 00046-1124246822.png)

(2.17 MB 1024x1536 00482-4150983202.png)

(1.87 MB 1024x1536 00057-4292960662.png)

(1.88 MB 1024x1536 00015-305506597.png)

(1.93 MB 1024x1536 00071-3945398186.png)

(2.10 MB 1024x1536 00021-1777880691.png)

>>17264 What's your vae dropdown selection set to? Does your vae file share a name with a model? Do you have your vae in the VAE folder instead of the Stable-diffusion folder?
>>17268 catbox for 2nd one?
>>17267 The VAE I want to use, because it doesn't, and it is. Like I said, it works fine in txt2img, but inpainting colors are mismatched. This might not even be a VAE issue, it just looks like one because everything I'm inpainting is desaturated compared to the original gen, but I don't know what else it would be.
Man I've been trying all the learning rates to get this small image LORA to be flexible with text2img outfits, what a pain
>>17264 >It's inpainting, as far as I know it's not even using an upscaler When you're inpainting the masked area only it is upscaling it, since it needs to contain the inpaint area into whatever your resolution + padding size is
>>17275 I'm not inpainting the masked area only.
>>17262 Usually 3/4th msrp then about half six months later
(596.71 KB 2912x2048 107513671_p7.jpg)

How to achieve a proper fake anime screenshot style
>>17278 by using/making a anime screenshot style lora? There are so many in different styles so you have to be a bit more clear what you want?
seems like lama cleaner got an extension with the webui https://github.com/aka7774/sd_lama_cleaner
>>17272 Yeah, probably not a vae issue, I sometimes get mismatched saturation too, sometimes it's even more saturated.
(1.38 MB 1024x1536 00024-1119826723ed.png)

>>17282 Dammit, the hand didn't look so fked up to me before I uploaded. Whatever
>>17280 Judging by the issues it's pretty buggy. Would be amazing to have lama built-in though
(1011.42 KB 1024x1280 catbox_mbthtu.png)

>>17272 try swapping out whatever dedicated inpainting model your using to just a normal model, and then set controlnet to use its own inpaint model which replaces the need for dedicated inpaint models
I still don't understand what is the fucking purpose of inpaint models, I inpaint with normal model all the time
Can someone help me? I updated lazyscript which seems to have a nice ui now but I cannot find things like resizing or merging lora... How am I supposed to tinker on my shit loras without these?
>>17286 I'm not using a dedicated inpainting model. Using CN inpainting has the exact same result. https://files.catbox.moe/g6y6ee.png This is literally just the word "loli" prompted in Based64 with no VAE altogether, dragged straight into inpainting over her hands. Nothing fancy. I do not understand.
>>17289 I wish I could help but I just wanted to say that I think that gen is very cute.
(1.51 MB 1024x1536 tmpwk6sowt7.png)

>>17292 Something I am working on right now Normally I would tinker a bit before uploading but that has to wait I think >>17291 https://files.catbox.moe/cigfqq.png https://files.catbox.moe/qa04u0.png https://files.catbox.moe/9iib6j.png https://files.catbox.moe/9iib6j.png
>>17280 It only adds a tab to the UI, doesn't even have any sort of integration like the Photopea extension does. There really should be an extension for embedding arbitrary sites into their own tab if people want that so badly. >>17287 Inpaint models/inpaint controlnets are trained on pairs, so it has a better understanding of coherency. Try inpainting a huge swath of area at full denoise strength with a normal model or without any controlnet.
baking a kidouko lora while i watch nightdive employees and their paid zoomers try to defend and shill their garbage system shock demake on /v/
>>17289 I'm sorry! I didn't have time to add it yet! I plan on trying to get it added by the end of this week. Along with block weight training. I have to mess with the structure of my code a bit more as I'm going to be adding the utils in its own tab. It will have all of the options within their own tabs in that tab probably. I need to figure out how to handle it. For now if you want, the old scripts are still available on the old_scripts branch. You can install that following the install instructions on that branch. https://github.com/derrian-distro/LoRA_Easy_Training_Scripts/tree/old-scripts
posting for posterity, it's an alternate version of one of my older LoRAs I never posted that gets a way more interesting style most of the time, though sometimes it ignores the prompt https://files.catbox.moe/mzqr9t.safetensors
>>17286 is this a model for controlnet that needs to be downloaded is it in the latest version? i havent pulled my webui or extensions for a while since getting burned by that gradio update
>>17299 ControlNet extension 1.1.x and the ControlNet 1.1 models. Lots of useful new features.
>>17290 Okay, so I turned off the "Ignore selected VAE for stable diffusion checkpoints that have their own .vae.pt next to them" setting, even though I do not have a single .vae.pt in my checkpoint folder, let alone one that's for the models I'm using, and that fixed it.. except, now, if I try to use ControlNet inpainting, that's cooking the fuck out of it. Whatever, at least I can inpaint at all now.
>>17301 I double checked the code for instances where opts.sd_vae_as_default is used (the setting you're referring to) and it's only used in one place, and if what you said is true it would have never mattered. If CN inpainting is now cooking it for no reason too your install sure is cursed. At least it (almost) works for you now??
>>17302 >your install sure is cursed Yeah, this thing's been through a hard drive failure and months of A1111 updates.. now that you mention it, I'm just gonna nuke the venv and start fresh. Hopefully that sorts me.
>>17303 It didn't. Whatever, I don't care enough, I rarely go to the effort of inpainting anyway.
(1.93 MB 1024x1536 tmp85j0b2xt.png)

>>17297 Thank you once again for your time to help out an useless anon Will do it with the old for now but I look forward the new, too.
>>17293 >>17305 Catbox pls? Very cute style.
(1.60 MB 1024x1536 tmpmfnrxpjv.png)

>>17306 will do when I am finished and upload the lora too currently installing the old scripts again
>>17307 Alright, thanks anon
>>17308 As it seems that I fucked up my pytorch by installing the old script somehow i will try to fix it first Here is the lora with a few examples before I could tinker with it https://anonfiles.com/M1sf34ubze/chahan_fried_rice0614_7z
>>17309 Will try it, thanks once again
(1.53 MB 1024x1536 catbox_t0noqj.png)

(1.46 MB 1024x1536 catbox_jlabh2.png)

(1.58 MB 1024x1536 catbox_mk24v6.png)

(1.57 MB 1024x1536 catbox_8xww8i.png)

>>17310 Fixed my fuck up Currently working on the V2 as I want to get a bit nearer to the original artists style https://gelbooru.com/index.php?page=post&s=list&tags=chahan_%28fried_rice0614%29+
>>17311 I like the first version too, thank you anon
it's a shame to see how fucking bad cuckchan hdg is now
>>17293 Thanks, but you have the third one there twice, can you post the last one again?
>>17315 https://files.catbox.moe/7upib2.png sorry normally I use the script for direct posting but sometimes it just refuses to work
>webui now saves images in a temp folder t-thanks
>>17316 Thanks.
(4.99 MB 3072x2048 tmpj39epoqa.png)

>>17312 >>17313 yeah while I like chibi the mix between normal and chibi style is kinda nice
(2.33 MB 1280x1600 catbox_jy5bvv.png)

God I fucking hate the schizo chinks on CIvitai, they love to wipe metadata info on LORAs and image posts
(1.67 MB 1024x1536 00007-859567552.png)

(1.72 MB 1024x1536 00035-788986431.png)

(1.86 MB 1024x1536 00050-921891856.png)

(1.67 MB 1024x1536 00005-3122086661.png)

(1.68 MB 1024x1536 00231-1869337264.png)

(1.72 MB 1024x1536 00290-146740118.png)

(1.70 MB 1024x1536 00301-3183774046.png)

(1.88 MB 1024x1536 00502-608937078.png)

>>17324 my biggest pet peeve on AI image sharing sites >originu promptu dunotu steel! >frigure out promptus & LORA on yo own yeah fuck you and your selfish ass too buddy
>>17331 embarrassing
(1.92 MB 1024x1536 catbox_lm9a2r.png)

(1.93 MB 1024x1536 catbox_6sy65m.png)

>>17327 Pretty interesting style, thanks. *slaps on 0.3 Zankuro*
>>17334 Oh yeah, Chewy marshmallow candy's a pretty good model, I like it a lot personally.
>>17334 I'm trying to find the Antifreez loras the models are using but I cannot navigate that site and search does not help
>>17337 there's one on civit and another I personally trained a while ago under a different name (probably worse in comparison) https://mega.nz/folder/oUokGaJR#OWRSRLgTrr9MrbPzkAFfYA/folder/oBhkUZjY
The pixiv shit is just reminding me how stupid private platforms are when it comes to avoiding the outrage crowd, well good thing most of us here are already using custom style mixes with LORAs and shit so that stuff doesn't effect me.
Is there some reliable way to get closed eyes that are arcing upward (when not smiling) instead of sleeping looking downward arc? I tried ^ ^ but that seems to affect only the mouth (smiling expression).
what the hell, the new update is giving me cuda errors when using hires fix at 2x? I'm confused because I have 24gigs
>he pulled
I’m going to seethe hard the fucking oldest commit I have used less VRAM than the newest ones, this is like commit a9fed7c3, Voldy what the fuck did you do
>>17349 Are you the same guy from cuckchan? I personally still am on a9fed7c3 on my local install
>>17348 Crop top + denim/bike/gym shorts is basically all I prompt tbh
>>17350 Does the latest controlnet work on it? Even my last gradio update install is still fucked with hires fix
>>17353 Don't know, I'm only using v1 locally as 1.1 rapes vram too hard. I've seen someone saying that it should work fine but I haven't tested it myself so try at your own risk
Isn't the VRAM situation due to the torch upgrade? You could simply use the older version of torch on the latest version. The TORCH_COMMAND env var exists for a reason.
>>17355 Pretty sure it's controlnet itself being heavier
>>17355 >>17356 I tried torch 2.0.0 and I still got VRAM errors on the latest commit, not sure what it is but I'm sticking with older commits for awhile.
>>17356 but your post might be leading onto something, it seems absurd but I was able to use hires fix before I installed all of my extensions again for the latest commit... I'm not sure if it's possible an extension had an update that fucked up vram usage but it's the only thing I can think of right now
>>17358 check if you have any extra shit enabled while generating?
>>17359 yeah I'll do a clean install again and see which extension is causing some weird shit to happen with the VRAM, the oldest commit I have has the basic extensions I needed compared to all the new shit I grabbed with the gradio update commits.
OK I confirmed my theory, there's an extension I installed that had an update that fucked over my VRAM when outputting, I'm going to find it to warn you guys but it's not the updated controlnet. I tested it out with the oldest commit I had and it didn't give me OOM errors, it has to be one of those dumb animation extensions I installed because I was bored and wanted to try it out, I'll post the name of the culprit if I find out which one it is.
>>17356 >>17357 I'm confused why it's been months and there really is no rhyme a reason for what the cause is. I have a hunch nobody seems to have bothered looking into it because those who would are either 1) already knowledgeable enough to figure out a workaround, so it was ultimately a non-issue for them (the "smarter" people are more likely using a lib like diffusers anyway), or 2) already have a 24GB GPU so it doesn't matter to them, which is the boat I'm in and why I've never bothered digging into it (sorry) I do think the tiled VAE as part of the multidiffusion extension solves almost everything if I'm not mistaken, but again this is a scenario where I've had no reason to bother testing because it's not an issue for me.
>>17362 Guess I should've listed a 3rd case where it's apparently just some isolated issue like >>17361 figured out, but that could also fall under the 1st.
>>17361 it's one of these extensions, most of the animation ones are shit anyways so yeah avoid them unless you plan on using temporal kit, that ones isn't so bad but it's basically "use ffmpeg and img2img batch your frames lole"
(1.81 MB 1024x1536 catbox_ozpzrd.png)

(1.97 MB 1024x1536 catbox_3l531l.png)

(1.97 MB 1024x1536 catbox_klovml.png)

(1.88 MB 1024x1536 catbox_vgt4ul.png)

I love this dragon
hey all, I finally updated the scripts. This update is a pretty big one, I added block weight training to the UI, as well as added back the tag file output into the UI. Next is going to be adding all of the various lora utilities, hopefully in a much nicer way than the popups. you can update using update.bat if you haven't already, or install using the instructions on my page: https://github.com/derrian-distro/LoRA_Easy_Training_Scripts
(2.22 MB 1280x1600 catbox_v9hsk3.png)

(1.88 MB 1280x1600 catbox_cd6nak.png)

(2.09 MB 1280x1600 catbox_ahkw8f.png)

(2.07 MB 1280x1600 catbox_toa0te.png)

>>17366 Thanks as always for the updates. Been using the GUI for a bit and liking how accessible the settings are and the overall polish. Not sure if I like individually selecting image subfolders but I like how you can change settings for each folder. For some reason the new update tripled my training time so I'm going to stick with the old popup for now for ease of queuing with json files. I'll definitely come back to the GUI when I feel like experimenting more.
>>17367 that sounds really odd. when I was testing I didn't see any noticeable reduction in speed. It's fine that you want to stick with the popups, just keep in mind that I don't really plan on adding to them anymore.
>>16584 bro #3 is clean enough to drink out of. got a catbox?
>>17366 hi, are your scripts configured to work with older 10xx and 16xx series cards yet? i haven't been able to use the default kohya_ss by maltais with my 6gb cards yet even though it should theoretically be possible. i just get black images or NaN precision errors
>>17349 how much extra vram is it taking for you and what card are you on? i'm able to use 2 controlnet models and a lora on a 6gb card with the newest commit on torch 2.0.1 cu117 and cnet 1.1. i could probably do a 3rd model and more loras, i just haven't tried it yet
>>17370 they have a patch for the 10x0 cards, and have had one for a while now. just make sure to install torch 1.12.1 not torch 2.0.0 or 2.0.1. You won't be able to install the triton patch nor the cudnn patch. so for you, once you use git to download the repo, run install.bat, then press 0, for torch 1.12.1 n, for the cudnn patch then y, for the 10x0 series fix. this should work for the 16x0 cards as well, I think. I couldn't tell you if you can really train with 6gb though, it's at the bare minimum i've seen work. and you can pretty much only use adamw8bit with xformers batch 1 512x512 training, possibly... I can't guarantee it working, but I can guarantee that you can install it.
>>17369 based retard
>>17366 thanks anon, just glad your updates don't break my shit like the webui desu, I'll post something if there's any issues with the current update
>>17374 please do, I try to test everything before I push it, but I'm only one person!
>>17366 Thanks anon, stay based
(1.77 MB 1280x1600 catbox_cyewoe.png)

(1.81 MB 1280x1600 catbox_iol3fp.png)

8 dim (9.4MB) vs 32 dim unet-only (26MB). One is not always better than the other but it's interesting to see the differences. Something like the mousepad lora just does not work as well as unet-only.
was there ever a good nazi uniform lora? i remember the one that got posted on /h/ forever ago and i also tried making my own but mine turned out worse than the other guy's
>>17378 nope, not yet also wanted to make my own but haven't gotten around to it yet
if anybody wants a good commit check before the gradio update ff216820fd680b5d04ce6cea7bf21e1bbec356b4 fixed issues I had with the additional networks extension
(2.31 MB 1280x1920 catbox_wokszc.png)

(2.10 MB 1280x1600 catbox_u90wc6.png)

(2.04 MB 1280x1600 catbox_7yu159.png)

(2.29 MB 1280x1920 catbox_6cjkej.png)

SSBU Character Render lora https://mega.nz/folder/2FlygZLC#ZsBLBv4Py3zLWHOejvF2EA/folder/vVdE0BqA Maybe not the best examples but I like how this one turned out
>>17382 Thanks >1024px for 8dim why
>>17383 1024x is the new 768x
I love these recent style tweaking LoRas. I don't know how useful all of them are, but it could definitely help you tweak a style in the right direction.
>>17385 Obviously, this is just a saturation lora. But there are ones for brightness, outline and line thickness, "face age", background flatness etc etc.
Sorry to put more work on you easy anon but if you see this Kohya has done really interesting updates recently >added an option --scale_v_pred_loss_like_noise_pred to scale v-prediction loss like noise prediction in each training script. By scaling the loss according to the time step, the weights of global noise prediction and local noise prediction become the same, and the improvement of details may be expected. >Max Norm Regularization is now available in train_network.py. PR #545 Thanks to AI-Casanova! Max Norm Regularization is a technique to stabilize network training by limiting the norm of network weights. It may be effective in suppressing overfitting of LoRA and improving stability when used with other LoRAs. See PR for details. >Three types of dropout have been added to train_network.py and LoRA network. Dropout is a technique to suppress overfitting and improve network performance by randomly setting some of the network outputs to 0 Wondering if you can add these soon, I want to test out if they really do have any improvements
>>17386 >>17385 Using LoRAs for basic color post processing seems like a waste when you could use an addon like https://github.com/Zuellni/stable-diffusion-webui-image-filters
>>17388 Thanks for the link! Obviously a lora is not needed for changing image saturation. There's a reason why I said >this is just a saturation lora There are ones way more useful like the popular detail tweaker or something like this eye size one. There's even one for easily making your gens more / less choco one might be interested in.
>>17388 >>17389 Here's one final example as I vaguely remember MaidAnon at one point talking about occasional difficulties creating his choco wife. This is with (dark skin:1.3) in the prompt and the mandatory 0.3 Zankuro
>>17390 I quickly learned what actually caused that, there's a few concept LoRAs that just can't do dark skinned characters but this could still be useful.
>>17389 Where can I find these LoRAs anyway?
>>17395 cool, thanks.
anyone know the perfect ratio between a main appearance folder and alternative outfit folder when it comes down to LORAs?
>>17256 >>17261 I should get around to finishing trails... Towa is so cute and erotic
>>17261 Can you share the lora used for this?
>>17387 I'll get them added today, I wanted to see what people did with them as well.
>>17398 I usually aim for a 1 to 1 ratio, because, to me, I'd like both (or more) outfits to be easily promptable. For example, my Arona + Plana lora had 1 to 1 for the base characters + outfit, and 1 to 1 for the swin skin to the base outfit, as such every folder had give or take 100 images in them, or repeats to make about 100 images, to match the highest number of images. Minus the folder in which Arona and Plana were together, I kept that at the 50ish images that were in that folder.
i can't trust using something like detail tweaker just for fear of limiting the latent space too much
>>17403 >>17395 meant that as a reply
>>17403 It's really good at neg values imo
(2.12 MB 1920x1280 catbox_1lk4ql.png)

I guess there hasn't been any updates on rescale cfg for kohya or A1111 https://arxiv.org/pdf/2305.08891 Haven't had much luck with noise offset but looks like I'm gonna have to keep trying
>>17406 it'd be nice if that was implemented, way too over my head to implement it myself.
(2.18 MB 1920x1280 catbox_77y0g5.png)

>>17407 Somebody implemented it for kohya gui but hasn't made a pr for it https://github.com/AMorporkian/kohya_ss ComfyUI also has an experimental node https://github.com/comfyanonymous/ComfyUI_experiments
>>17408 I see, I might look into modding sd-scripts to put it into my UI then, provided I can figure out how best to handle updating sd-scripts down the line. I also might just end up lifting their code for just sd-scripts and PR it if it works without too much tinkering, because I really still have less understanding than I'd like over how all of this works on an implementation level. but guess that's because I've spent so much time working on a UI that aims to beat out kohya_ss. I really do need to take some time one weekend to just deep dive into a bunch of papers to better understand all of this.
>>17387 hey all, pushed an update to dev that adds the new args. I just need to update the readme before I push it over to main, stay tuned I'll post when that's done
>>17410 https://github.com/derrian-distro/LoRA_Easy_Training_Scripts alright pushed to main, you can update via update.bat or install by the instructions
>>17411 thanks anon i'll test it out
(2.94 MB 1280x1920 catbox_wriavm.png)

(3.03 MB 1280x1920 catbox_fg95a1.png)

(2.88 MB 1280x1920 catbox_6pqg73.png)

(2.43 MB 1024x1536 ComfyUI_11868_.png)

>900k view video of some chink normie making an onlyfans with SD >uses LORA >I just kept training the LORA on the one girl I wanted with generated images >100 repeats method used I'm honestly hoping open source AI get nuked from easier access just because of this shit
>>17418 >nuke the tools, not the normies Wrong attitude
>>17419 I meant the methods of using/finding it easily, it's bad advice videos made by normies why the internet has just been plastered with shitty AI posts.
>>17421 That's pretty much censorship though. If you're gonna use a kike tactic I think you'd be better off spreading misinformation about hardware requirement and whatnot.
I feel bad leaving all these old LoRAs to rot, so here's a bunch of LoRAs that I can't post on civitai. Mostly loli artists. Sorry they are mostly 128dim because I trained them a long time ago: https://anonfiles.com/90v5o6v6z3/Unshared_zip
>>17424 Can be easily resized to 64, it's no big deal
>>17424 Pretty sure you can just prompt hags with the lora and post on civit anyway
>>17334 Started messing around with it a bit I have honestly no idea what im doing but its a start
>>17424 Can you upload to a different host? This site is shitting itself for me https://fileditch.com/ try this one for example
>>17426 You totally can, I've done it with a few of my style lora. Though when I posted utatanelolicoco I had to go out of my way to not use their whole name, because loli is in their name and I would have been taken down 100%
Anyone plan on making de-homofication LORAs for this special month?
>>17426 That was the plan, but from some testing the only ones I was ok doing that with were Lasto and Anyannko. If the Lasto one lives for a while maybe I'll add more, but people will recognize the more popular creators which is concerning.
>>17428 This should be better: https://pixeldrain.com/u/H6BtDu1N Also, for anyone that wants to know the contents in advance it's 19 art styles, 4 concepts, 2 characters, and 1 outfit. Some of them already have LoRAs made by others, but no harm in sharing these I guess. Also, highly suggest a weight of 0.8 for most of these
>>17432 >0.8 eh still better than most LORAs posted and I use style LORAs at that strength with character LORAs so good enough
(2.62 MB 1280x1920 catbox_iwl4aq.png)

(2.09 MB 1280x1600 catbox_h0njav.png)

(2.42 MB 1280x1920 catbox_lzrsbq.png)

(2.58 MB 1280x1920 catbox_79bc3c.png)

(2.44 MB 1280x1920 catbox_zssv6s.png)

(2.50 MB 1280x1920 catbox_qoxhwe.png)

>>17434 Was using an hll4 mix so the vagene got a little fucked on the examples
>>17424 >Sorry they are mostly 128dim because I trained them a long time ago 128 is no longer the meta? fug I just got started on LORA training and now its all wrong.
>>17437 it's the big difference between this board and 4chan, I personally believe that there is little reason to train at such a high dim, I usually train characters in dim16 and styles in dim16 conv dim 8, you can usually resize them without issue though.
(2.54 MB 1280x1920 catbox_sldwl1.png)

(1.89 MB 1280x1600 catbox_w5yuja.png)

(2.41 MB 1280x1920 catbox_wg34v5.png)

Updated Jahy https://mega.nz/folder/2FlygZLC#ZsBLBv4Py3zLWHOejvF2EA/folder/bE8kBQbD Going through and redoing my Character loras so I'll probably just post again once they're all done.
>>17440 what are you redoing? All the new options that were added with training recently?
>>17439 huh, well alright guess I shouldn't follow the now outdated guides to a "T" then. (excuse the /aco/ everything else ive trained has been non-cunny)
(2.95 MB 2400x5918 catbox_u6a2bo.jpg)

(990.79 KB 640x960 catbox_5gsp88.png)

>>17441 In some ways my old ones are better than my current ones but it's mostly just to make them low dim. I also used to use reg images for a character loras which tended to have odd effects like ignoring prompts or adding in random background elements
Is there a guide somewhere for regional prompting?
(2.25 MB 1600x1280 catbox_rcv2ze.png)

When your feet are in the water for too long
(1.91 MB 1600x1280 catbox_qod9u2.png)

(6.76 MB 2048x3072 catbox_8thcxy.png)

(1.58 MB 1024x1536 catbox_rop25s.png)

(1.28 MB 1024x1536 catbox_bz6ehx.png)

(1.95 MB 1024x1536 catbox_bkuol5.png)

been a hot minute since I've posted anything, haven't really been able to gen, but here are some gens I made the last two days had to get a gen in with hifumi, because I haven't been able to lately.
(1.17 MB 242x243 feet.gif)

>>17442 it really doesn't matter unless you're concerned about large size of 128dim
>>17449 large file size isn't the only thing that higher dims can be a problem with. much more prone to learning unwanted "noise", especially given that 99% of the time, dim128 has like half of it's space not even being utilized. regardless, just my reason why I don't use dim128 lora
>>17450 > much more prone to learning unwanted "noise" in theory sure. in practice, i haven't found something 128dim couldn't do that lower dims could. even then, resizing or simply training with a lower alpha will fix those issues. i appreciate the lower filesize, but idc what dim someone uses as long as it's decently trained
>>17451 >idc what dim someone uses as long as it's decently trained yeah, specfagging over ai params is pretty kuru kuru kururinge
>>17451 I dunno, i've seen better results with lower dims. so I don't download dim128 lora, unless it's something I really like, in which case I'll resize it, if possible
>>17453 the real question I have is... why don't more people train locon? they generally allow you much more freedom in training at lower dims, and as an added bonus learn styles better.
>>17444 The github page has some pretty good examples imo, have you looked there?
>>17454 Probably because there aren't any tutorials for these, or at least I haven't seen one, but that might be because I don't really pay much attention to ML except checking these threads from time to time these days. As for myself, I still have no idea what these post loras actually are or how to train them, and I'm pretty satisfied with just loras, so I don't bother.
>>17440 >1e-3 huh?
>>17457 e3 works best for 1 alpha when it comes down to adamw
apparently the upgrade to d-adaption (3.0) is using way too much VRAM and it's not worth it, you can pip uninstall d-adaption and pip install d-adaption 2.0
(1.45 MB 1024x1536 00052-3163032120.png)

(1.60 MB 1024x1536 00295-2886003196.png)

(1.51 MB 1024x1536 00020-2643177040.png)

(1.81 MB 1024x1536 00572-146740118.png)

(1.42 MB 1024x1536 00026-766230012.png)

(1.59 MB 1024x1536 00026-3693188573.png)

(1.60 MB 1024x1536 00024-1965698000.png)

(1.80 MB 1024x1536 00041-13876563.png)

>try working on double outfit LORA >still having trouble skill issue or I really have to focus on manual tagging ugh
I don't know why people have issues with the easy GUI, heck it's gotten better since the first release and now that the queue option is fixed I'm using it mainly now. I don't know if some parameters make it different from using a powershell script with the main repo but at this point I'm just trying to bake shit
>>17454 I've never found that using locon improves anything and actually just ends up taking up more space than just using higher dim, I could be wrong so I'm working on some more training samples.
Not sure if the new args that Kohya has added to the scripts are doing anything "good", have people tested out the new stuff? I'm just too busy trying to get this multi outfit character LORA to work correctly to focus on testing.
>>17468 I'm retraining with more of the new parameters right now but here's the effect of network dropout Left is no dropout, right is with `network_dropout=0.1`
>>17456 you got a point, I intended to have a tutorial out already for my training style to go alongside the release of the UI, that got delayed because I ended up just working on the UI some more. I really should get around to formatting and publishing it so that there is examples of how to train LoCon out there, because they really are useful. >>17465 thank you! I'm testing out a bunch of models to see if I can create a unique style, those were all just gens I made during that. >>17466 I could have sworn I made the UI really intuitive and easy to use. I am also surprised to hear that people are struggling, especially when I haven't gotten any real bug reports on the github. >>17467 I've definitely found that LoCon shine in compatibility. I think because they train on more layers the weights are more even which means the weights are not all bunched up on a few layers, meaning you can use them at strength 2 and not see any frying, it also means you can use two or more at 1 together and not see frying. I also find that it does produce much better results with smaller file sizes, but that might be because of the way I train. I don't usually go higher than dim32 conv dim32 for locon I make, usually getting into dim16/alpha8 conv dim8/alpha 1. this usually results in a very close fitting style LoCon that is like 28mb in size. sometimes I do need to go higher, but even then it's more or less usually less file size than the equivalent lora would be, in my testing that is. Like I said above though, I really need to get around to posting the updated guide which explains things like style lora training guidelines among other things. >>17468 A friend of mine did some testing on them and found that a lot of the recommended args aren't actually good, but did find that you can use these settings to seemingly better encourage the lora to train on as many layers as possible as evenly as possible: network_dropout = 0.5 min_snr_gamma = 10 scale_weight_norms = 5.0 the other args seem to be less impressive.
>>17469 the only positive thing I notice is her skin is more dark, I guess that's good if the character is tanned and her gear on the side makes more sense appearance wise, it's really minor difference wise desu >>17470 That snr value is high but I'll test it out, so many values but I guess most of them don't have such a huge effect huh
>>17471 seems like for anime style training an snr of 5 actually can do more harm than good. so by having the value be higher (thus weakening it) it will ignore less of the smaller details
>>17472 would you have any comparisons? curious about this since i've no idea what the current meta is
>>17473 I don't unfortunately, no. I'm mainly reporting what my friend has told me, I haven't really used the args that much so I don't really have anything to use as basis.
(603.93 KB 2400x2539 catbox_3uxavp.jpg)

Not sure what to make of this
(545.57 KB 2400x2539 catbox_1zdi1z.jpg)

(1.94 MB 1024x1536 catbox_mb38z6.png)

(1.91 MB 1024x1536 catbox_43p01b.png)

(1.90 MB 1024x1536 catbox_br5j1y.png)

I recently remembered about that bluesteel haired loli dragon girl I genned a while ago and decided to gen some more of her, this time in a maid outfit. I didn't upscale them to ludicrous levels like I did the previous time I posted on here, but they are still pretty nice. I didn't use any style lora this time, the base model was enough to carry it to be honest.
>>17477 >maid outfit nurse outfit, I must not be thinking, fapping too much I think
Minigunny
>>17478 >nurse outfit That's something I haven't genned in forever.
>>17480 same tbh, I just wanted to for some reason
(819.11 KB 960x1440 00361-979789203.webp)

Tried a thing
>>17482 turned out pretty good all things considered. nice job!
(2.60 MB 1280x1920 catbox_cw0ffb.png)

>>17483 Thanks, definitely expected it to be a lot worse since it was just controlnet and using the same seed/prompt. Even so it's funny how massively the style can change. Here's the source template https://files.catbox.moe/mbfwoz.7z
>>17484 oh cool! I might mess around with it, if I have some time. I still got a lot I want to add to the UI, so that's still most of my free time lel.
(2.02 MB 1280x1600 catbox_97qojb.png)

uoh
(2.16 MB 1280x1600 catbox_l84u23.png)

>>17486 uoh 2
>>17489 good luck to the (((yakuza))), i'm behind 3928 proxies
(1.26 MB 1024x1536 00469-3373191766.png)

>>17489 Every time someone links twatter I fucking know I'm up for the most abhorrent and stupid shit ever and I still am surprised I wonder if the troons are going to shame nvidia for cuda cores any time soon. Allowing local use of AI is raycis n shiet ya know
>>17491 >raycis We can use rays for our gens now? I can finally put those raytracing capabilities to good use
(535.46 KB 583x568 msedge_QLsU7k62wo.png)

ABSOLUTELY TALMUDIC
>>17491 >>17493 kikes and trannies need to be hung, good thing more people are starting to push back against trannies now so I'm hoping more of them are hung in the future to spare us from this retardation
>>17491 >shame nvidia for cuda cores It's gonna be so fucking funny, not even tech trannies understand that CUDA cores ARE the GPU.
Built in openpose editor for the official controlnet addon that lets you edit the preprocessor output: https://github.com/Mikubill/sd-webui-controlnet/discussions/1548 Could be cool for doing quick pose/hand/face fixes I guess.
>>17489 wow! im a wanted felon! (in Japan)
>>17489 So this does this shit have legal power already or not
>>17498 >>17499 Could just get companies like pixiv to pussy out which would be annoying because then I would have to find a loli friendly platform
>>17500 ugh reading the way this faggot types hurts my head but basically it's common sense copyright law, niggers can sue you if you do AI outputs that are not derivative of existing works and promoting it in a commercial sense, pixiv has already banned that and they've gone further to target accounts uploading derivative works with style/artist LORAs so it's not such a huge hit on that front.
hate this faggot but it seems like the recent Jewvidia driver updates are causing slowdowns when all the VRAM is being used because to avoid OOM errors it just goes into sharing memory with your PC's RAM https://github.com/vladmandic/automatic/discussions/1285
>>17499 >>17498 Isn't it just for COMMERCIAL use? Like if you're making shit for yourself then cool but if you make AI art that infringes on someone's IP and try to sell it then they can sue?
>>17504 Some people here seemed to do commission money. I personally do this for fun/fap so yeah it doesn't affect me yet. Would suck to see AI completely banned from everywhere because of angry niggers from twatter
>>17505 >because of angry niggers from twatter oh you sweet summer child
>>17497 Very happy with this. I was using third party extensions and while they were fine, the openpose hand implementation sucked.
>>17509 >redid teto. YEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEES >synthv outfit is pretty gatcha ...oh :(
>>17510 also thanks for reminding me i need to deactivate her and synthv before i wipe down the ssd and move it to a new build
Any tips for merging models? I kind of want to merge the chewy marshmallow model above with AOM3 or something else. Or has anyone done model mixes with it so I don't have to?
>>17510 it just tends to flub the details more than half the time and has no idea how to recreate the striped "staff" pattern on her uniform. i think a lot of the posts weren't tagged properly looking them over again.
>>17502 That would explain the slowdowns I've been having lately. Funny how just a week ago they promised 2x Stable Diffusion performance https://blogs.nvidia.com/blog/2023/05/23/microsoft-build-nvidia-ai-windows-rtx/
>>17514 Did anyone even benefit from that?
Might want to get to baking that new thread
>>17515 The "speedup" requires converting pytorch models to onnx. You can convert existing models but I haven't bothered to try because it's also not compatible with lora https://devblogs.microsoft.com/directx/optimize-directml-performance-with-olive/


Forms
Delete
Report
Quick Reply