was about to hit post >>19951 I hope so the 120 pages of fat gura won't host themselves

I have migrated

>>19954 The extra noise parameter I added is too kino. Those extra streaks of water on her body...

>>19955 Damn, it does look pretty nice, and doesn't seem to fuck up anything.

controlnet is hard

>>19957 you're getting there champ

https://www.pixiv.net/en/users/54698934 Anyone have a LORA of this artist or did anyone ever make one?

one more time, I promise, just testing the catbox script for something stupid https://files.catbox.moe/uq4g2v.png

>>19962 alright, didn't seem to work, but if you wanna see what I was doing, drag the image into PNG info

>>19959 Not mine but here's a higegepon lora https://files.catbox.moe/f4f709.safetensors

>>19964 Based, going to try it out later.

THE ABSOLUTE STATE OF LTT SHILLS LMFAO

>>19966 Unironic coomsoomer brainrot

>>19967 >n-noooooo you're harassing a poor millionaire!11!!

>>19954 >migrated pls explain I'm out of the loop

>>19966 kek

the amount of seethe that sd still is able to milk from normalfags is astounding man. even on freecodecamp video there's faggots screeching about muh starving artists and stealing. I wonder how this whole shit goes long term

did you invite /g/ in here or why has this place become so cancerous also >anonfiles talk last thread that site was hosting tons of CP since TORfags were using it

>>19972 When you get tired of watching sd/g/ you kind of want to check out what else is going on. And yea, anonfiles was being used to host questionable shît. Even the forbidden model was on there at one point.

>>19972 This thread is an offshoot of an /h/ thread which was an offshoot of a /g/ thread, I think most of the posters here originate from /g/.

>>19972 Some cunt probably linked it on /g/ or some faggot discord. >>19974 People that currently reside in /sdg/ general are mostly completely braindead niggers and pajeets, it wasn't nearly as bad few months ago.

>>19955 So what is it exactly? I read the PR and it seems like CN tile.

Why do people keep training anime on SDXL

>>19977 civitai contest

>>19977 >>19978 pretty much this, just sunk cost and sour grapes

>>19976 Well, I made the PR and that's not really what it is. That being said you actually could use this with CN tile if you want to, it might actually improve it. https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#extra-noise I also made a PR and script for masking the area it adds extra noise to (for hires fix). Later I was going to see if I could hack it together with DAAM or however Regional Prompter calculates the attention area so you could just prompt which area you want to be masked, but I honestly still prefer doing it manually. https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12616

>>19980 Doing gods work anon!

>>19980 did you have any prior ML/torch knowledge before sd? can someone who knows somewhat basic flask python learn and contribute in reasonable time frame?

>>19982 I had none other than screwing with ML projects for years just for fun (GPT2 days). Also I frankly still only have surface level knowledge of SD. If you look at the PR that implemented the extra noise it's actually incredibly simple. Anything outside the pytorch stuff shouldn't be too difficult to pick up, depends what you'd want to contribute I guess.

>>19983 I see, thanks. >depends what you'd want to contribute I guess Dunno. I kinda have too plenty of free time currently and most likely will for a few months, so I just thought it'd be cool to do something useful with this.

I am going to try looking into the NovelAI VAE issue more, because it seems like it only happens in hires fix, not img2img, which logistically doesn't make sense because they're the same operations. I first want to make sure it's consistent across systems though. Can someone confirm they can generate this image fine, but using hires fix causes the NaN issue, even with the VAE in full precision (--no-half-vae)? The only hires fix setting that seems to matter is upscale factor, it starts failing at 1.15 for me. --no-half actually fixes it entirely but the whole model shouldn't need to be using full precision, might be a clue to the issue though. https://files.catbox.moe/74wkie.png

>>19985 Would be easier to help check if you catboxed one that doesn't use two loras.

>>19986 If you can make an example without LoRAs then please do, that would help. Here's the links incase you need them: zankuro https://litter.catbox.moe/g9rgm1.safetensors ponsuke https://litter.catbox.moe/h2b8wj.safetensors

>>19985 >>19987 Also I think I already am making some good progress, it seems like it's something going wrong with the image that gets fed in for conditioning in img2img, not the latent

Is there no latent upscaler apart from basic shit? How difficult would it be to train one

>>19985 Works fine for me. >>19990 >but I can't find the setting that enables the inclusion of VAE hashes in the metadata. That's one is still in the dev branch iirc.

>>19991 Forgot to mention I use --no-half-vae.

>>19990 >>19991 Well shit. Yeah, that's the correct one. I just used the metadata from those upscales as a sanity check too and it still fails for me. What's more weird now is I don't think it's the latent or the image conditioning necessarily. It's consistently failing within the decoder upscaling, specifically here: https://github.com/Stability-AI/stablediffusion/blob/cf1d67a6fd5ea1aa600c4df58e5b47da45f6bdbf/ldm/modules/diffusionmodules/model.py#L634-L641 I had more to write here but I might actually be onto something now, hopefully I have a better idea of what's happening soon

>>19925 you living the dream, anon. just make sure there are no reflective surfaces behind you on those conference calls.

>>19993 Can you post/verify the vae sha1 hash? I got the one from the old torrent like >>19990, same hash too.

>>19994 Kek

>>19995 Well shit, you did verify, I'm blind.

>>19985 Will verify later but I get NaNs even on normal gens, is that fixable without forcing me to use nohalf? Cause that one doesn't work either

>>19990 >>19991 I think I narrowed down the issue -- if one of you are willing, could you checkout the dev branch and see if it works or not? If not, how you could help instead is find some prompt/setup on 1.5.1 that produces NaNs. Maybe it was somehow completely resolved as of 1.5.1 though and the upcoming 1.6.0 reintroduced it. Specifically after doing a git bisect this is the commit that reintroduces the issue for me: https://github.com/AUTOMATIC1111/stable-diffusion-webui/commit/cc53db6652b11e6f7bca42c3aa93bd6761ed3d3f

>>19998 Catbox one of your gens and I can check. Ideally it should have no reliance on 3rd party extensions (built-in are fine). Will need commit hash too.

>>20000 Will send one later. I'm on an old March commit but stopped using the NAI VAE in January cause the NaN errors were frequent enough to bother me even with nohalf and even when generating without hires/img2img

I'm on 22bcc7be still, anything I can do to help?

>>20002 Make some prompt setup using only built-in features that causes NaNs with the NovelAI VAE even with --no-half-vae. Using --no-half fixes it entirely in my tests but of course that's stupid to make that necessary.

>>20002 >>20004 Actually, if you can do it even without --no-half-vae, that would be even better, but you need to confirm then if it happens in img2img. You can make img2img behave identical to txt2img if you set it up like picrel. If you're on 1.5.0 or above you should disable the "Automaticlly revert VAE to 32-bit floats" setting too, otherwise it forces the webui to use full precision until it's restarted.

>>20005 >>20004 I do not have the automatic vae setting and never have --no-half-vae on. So make a gen until it gets the NaN error, then try it again in img2img?

>>20006 Yeah your version is hella old so it won't have that. >So make a gen until it gets the NaN error, then try it again in img2img? Yes. If it still produces NaNs in img2img then that very likely indicates it's the VAE and nothing can be done. If it doesn't then something is wrong with txt2img.

>>20007 I'll get to work, This may be some helpful info, but one thing I noticed is that when I start doing img2img work, whether upscaling or inpainting, NAI.vae will start acting up and shit out NaN errors almost consistently and I then have to switch vae.

>>20008 Ok, so I managed to get an image that NaN'd out in t2i, and then I recreated it on img2img with the settings you provided and it generated just fine, should I give you anything?

>>20009 Catbox the txt2img one (just switch out the VAE temporarily so it can actually output the image)

>>20010 https://litter.catbox.moe/71pmak.png - t2i (NaN'd Black Box for PNGInfo) https://litter.catbox.moe/eibubq.png - i2i https://litter.catbox.moe/a2xonj.png - t2i (Kl-f80anime2.vae.pt)

>>20011 oh wait nevermind, I just realized I did the img2img wrong and left the resolution at 512x512 instead of doubling it to 1024x1024

when I recreate it in img2img I get a completely different result

>>20010 Ok, so not sure if this would be helpful but I managed to generate an image that NaN'd without hires on NAI.vae, and attempting to recreate it in i2i inpaint also NaN'd/Black Squared https://files.catbox.moe/s2qi6h.png original t2i https://files.catbox.moe/ns0vxp.png kl-f8-anime2 https://files.catbox.moe/7bvs8m.png original i2i So I take it this may not be good news

>>20011 >>20012 >>20013 Thanks, that's helpful, I was able to reproduce the NaNs too. I actually forgot to mention in >>20004 what I really want is an image that produces NaNs _without_ hires fix, because this seems strictly limited to hires fix so far. Is your exact commit 22bcc7be like you mentioned? If that's the case I'm gonna have to do another git bisect because it seems like voldy fixed it at some point but then fucked it up again. Reason being is that it works on the release commit of 1.5.1 and now my PR's attempt to fix it (https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12624), but does NOT work on the dev branch and (presumably) the very old version you're on. >>20014 Yeah this is what I really wanted, cheers. I can confirm that one fails in fp16 but works in fp32. This gives me enough to start looking into a possible root cause.

>>20015 yea it is

>>20015 Maybe the last thing that would be useful is an image generated without any special prompt syntax (including weights) and without LoRAs, and on the base NAI model. Reason being is then I could compare with Naifu when it's in fp16 mode and see if it craps out or not.

>>20017 Doers special syntax include BREAK? Because I just got a black square after removing everything else

>>20018 >Doers special syntax include BREAK? yes because NAI never gave the user the ability to delimit the token chunking like that

>>20019 ok, prompt is currently >scan, traditional media, sketch, extremely detailed CG unity 8k wallpaper, hakurei reimu >worst quality, low quality everything is on euler a defaults no hires, gonna run generate forever and see if I get anything

>>20020 bless u, anon I have a hunch NAI just forced fp32 on their servers to work around this but I'm hoping that isn't the case. If it is then they literally did just do some DK64 expansion pak type shit to work around it.

>>20021 it's been almost 10 minutes and 300 images later and I haven't gotten a single black box/NaN'd error. I think these special syntaxes are part of the problem

>>20022 (nta) it could be some numerical error when adding two different prompt chunks (idk what's the actual name, not deep enough in this) where a "zero" value works but zero + zero != zero and this blows up somewhere this is probably just schizothoughts though

>>20023 you mean two different 75 token chunks?

>>20022 You're free to try weighting stuff then, I'll just have to double check the weights end up matching in Naifu. Anything else is basically off the table though.

>>20024 yes, since that's what BREAK is supposed to do

>>20023 It won't be this because the part where it breaks is upsampling the latent -- tokens don't matter at that point. It does influence the latent result of course but the numerical error is occurring way after tokens are handled

>>20025 yea lets do it, let me know what you need

Apparently the sharding is also usable on inference, so multi-gpu/tpu inference with low vram (they say 8/12gb) for SDXL seems possible So you don't need a giga gpu, but multiple smaller ones seem to work

>>19999 Do you still need this? I don't want to copy a whole install again for nothing.

>>20030 nah I figured it out for the dev branch, at least for now. all that remains is 1) is there any simple prompt without LoRAs on base NAI that produces NaNs while using fp16, so I can compare to Naifu, and 2) if there are any that somehow still produce NaNs on fp32, if casting the VAE to bf16 fixes it (this doesn't work on dev currently but that is for unrelated reasons)

hdg is under comfynigger attack

>>20031 >(this doesn't work on dev currently but that is for unrelated reasons) meant to specify this is for fp32. I know bf16 requires Torch nightly and I have tested it before to work, just not sure if the NAI VAE has any issues with fp32 or not.

>>20032 seems to have calmed down, shill is probably on lunch break

>>20032 and I apologize I'm cluttering this thread with technical VAE shit so here's a pic

>>20035 It's all good, I gain a lot from having a fixed NAI.vae platform so I'm happy to help with any more things.

>>20034 they probably gonna try grooming people in other generals, wonder if we're gonna see those niggers pop up in vtai and trash lol. >>20035 where would you even talk about sd tech? this thread is perfect for this. Thematically sdg would be the best but it was overtaken by comfynigger. also cute Marisa

>>20037 There is at least one person on /vtai/ using comfy but they haven't done anything disrespectful (yet) so I've held back from calling them out

>>20038 yeah just using it is certainly fine, but fuck comfygroomers

>>20031 Well I was already halfway through the setup before you posted so I did it anyway and it does indeed produce NaN errors with or without --no-half-vae (actually a black image without and straight up nothing with).

>>20037 >also cute Marisa ty. I stopped visiting /sdg/ around December or January, now I just open the thread to entertain myself briefly for a few minutes every week or so. Looking back I think I really only enjoyed it until about a week after the NAI leak settled. /hdg/ started trailing off after December too but at least it isn't like /sdg/ yet. >>20038 I don't even watch vtubers but /vtai/ is consistently cozy, dunno how they do it. Granted I'm rarely in those threads but anytime I take a peek they're sane

>>20040 Okay, that part is consistent then, thanks for verifying. When voldy is online later he might know what's going on. He did some shit to support SDXL with hires fix but I can't tell why he did it the way he did.

>>20035 you're doing god's work, anon. please continue

>>20041 /vtai/ is the only comfy place on /vt/, the rest of the board is a schizophrenic psychological battlefield (they make neo-/sdg/ look like a joke), but the trouble makers want maximum coverage so they avoid a small coom hangout like /vtai/ and instead go after the global generals or make attention grabbing threads to trigger fanbases.

???

>>20041 >>20044 yeah I don't care about vtubers in general but the vtai is nice. I really wish there was a good anime general suitable for ecchi/sfw posting just like vtai.

>>20046 /e/ is the closest thing to that unfortunately, /a/ would shit a brick if we tried making a rotating AI thread

is it a good time to pull

>>20047 /e/ is absolutely dead. a would be nice, yeah, but their artnigger circlejerk is certainly not gonna let this work

>>20045 might just be an autistic anon that watches PRs like a hawk (like me) >>20046 >>20047 /e/ is fine and I don't even mind it being slow but there's a lot of newfags compared to /vtai/

So what are we looking at now with the vae testing? I'm pretty much from all day other than doing image crops that I can take a break from anytime.

>>20051 (Me) >pretty much free all day***

>>20051 >>20052 see >>20031, maybe skim the thread a bit too tl;dr just seeing if novelai vae is repairable or not

Finally found a nice tool to edit png metadata without doing exif shit or using inpaint with no mask at 0 denoise to insert whatever I want. It's called tweakpng. It also has an option to add it the context menu, so editing is fast and easy.

>>20054 damn I should've just mentioned this months ago. I've had it installed for years lol

>>20056 Extremely fucking based, on my todo list now

>>20056 >>20057 Lmao

>>20058 >>20057 >>20056 kek it was my homura pic guess that guy is gonna get a wojak instead of catbox

>>20061 >>20061 KEK

>>20060 You can also copy paste a tEXt block from an image that has metadata and edit.

be careful, that SAI troll sometimes stalks this place, he may tattle on us kek

can't wait for you fags to start crying about how bad this place has become once you're done stepping in shit on /g/ and then tracking it all over the carpet here

>>20065 womp womp go back

>>20066 zoomie doesn't understand what goes into building and maintaining worthwhile places, treats nice things as a toy, ends up with nothing and cries about how he's been done wrong by the world. many such cases.

I may actually install cumfy in order to create a "workflow" that is just boxes making "NIGGERS" to paste that in catboxes stealth

>>20068 just copy one of their text parameters and shit it up so it crashes

>>20067 couple of catboxes with hanging troonjaks won't force sainiggers here but yeah I agree that getting involved with that shithole too much is a bad idea

ill scrub my shit but im gonna still do it

>>20069 Are there any nodes capable of changing/moving/deleting files? May make for useful shenanigans >>20070 too bad the hanging troonjaks are too wide usually to render properly at least the last one I posted seem to render properly also got trips

>>20067 based boomer

>>20067 cute

>>20067 I've been enjoying this lora

So, is the NAI VAE fixable? Back when I used it I'd get NANs on both normal gens and hiresfix - with or without --no-half-vae. It was more common with hiresfix but it also depended on which upscaler I wanted to use, sometimes when Latent refused to work Anime6B would actually go through.

>>20077 If you have one that fails even with --no-half-vae I'd like to see that because it might work with bf16. Don't use the dev branch currently because there is an issue that makes it more likely (turned out to be some pytorch autocast issue)

>>20078 Whats the implication of getting a NaN error on a no-half-vae? How does that help with identifying a fix?

>>20067 this place has already gone to shit, gj whoever was retarded enough to link on /g/ and invite the current crowd. Ruined it

>>20080 are you lost?

>>20081 aren't you?

>>20080 This place has never been linked on /g/ The schizo shitting up /h/ that sometimes comes here found the link in gayshit

>>20082 Been here since #2.

I was just gonna ask how long others have been here for and then realized the old threads are gone. Gonna have to trust my memory but I think this is the same image I posted for the original hdg ban waiting room thread.

>>20084 newfag.

I’ll have you know I figured out the original riddle on my own before someone practically spoiled the Mario game and I bitched about it lmao

>>20035 catbox?

>use twatter once a year >need to DM someone >open twatter >find this nigger on my timeline (idk who he is) Do arthoes really? How can they claim they're starving when they dare charging so much for so little?

>>20089 it's just a tired theme at this point man. yeah faggots are ready to spend hundreds of dollars for a doodle just because it was made by their idolized artist. they pay for the name and feeling "honored". same reason why drawnigger threads still exist. they don't want to see an art of their waifu, they want attention from some cool discord kid.

Anyone got miyamoto issa lora? I feel like I've seen one somewhere but can no longer find it

your gay little drink, sir

>>20093 Very nice tan. Catbox pls?

comfyUI question: Let's say I wanted to do something like, e.g, changing an input to the sampler at every step. It's not what I'm doing, but let's say I wanted to change the cfg every step, oscillating between 1 and 5 like so: 1 3 5 3 1 3 5 3 1 3 5 3 1 etc until the gen finishes. At present the only way I know how to do this is to have as many advanced sampler nodes as there are steps. Since that's obviously not the right way, what is the right way?

>>20095 >he fell for the meme if you actually want to do advanced shit like this you're better off just using python directly

retard-tier question, does comfyUI by default deny requests from non-localhost? I'm trying to connect from another machine to the web interface and I can't, I've checked that my firewall is disabled. Also, is there a way to change the default port?

go back

What the fuck is happening today?

>>20085 I posted >>2 RIP

>>20093 Hot >>20099 Full pajeet offensive

What happened to the board owner? Did he die?

>>20102 if he is then I'm just going to post furry again

>>20103 Please don't.

>>20104 you WILL look at (spoilered) furry holos when I decide to tune this shit lora again until I get it working (I may even try training it on vpred zSNR furry models and see how backwards compatible it is)

>>20015 Black blank gen is pretty much the NaN errored gen, right? I just got one with no-half-vae. I tried using different prompt for upscale pass and it failed.

>>20106 yes it is >not catbox anon, but the guy helping him yesterday

>>20107 Damn I got 7 loras used there lol. I feel like it blanked out because I used different hires fix prompts, it didn't happen before. I'll do some more though.

>>20108 >>20031 >2) if there are any that somehow still produce NaNs on fp32 I think he still wants that

>>20106 >>20109 Yeah, I wanted to see if a failed fp32 gen works in bf16. The fact it's using LoRAs and such makes it less useful but I'll take what I can get

>>20111 I haven't tried doing a no-half-vae attempt but I didn't get any luck generating an image without a NaN error when not using special syntax running it normally. Should I try doing a run on no half but with weights and BREAKs only and see what I get?

>>20111 catbox keeps shitting itself so i cant upload the loras now. sry Im going to sleep will post tomorrow hopefully

After I finish these images sets I need to crop I'll do a fp32 no-half-vae run without LoRAs and see what I can get

>>20112 >>20113 Do you mean with a NaN error? Either way no rush anon, rest well

>>20115 Word, cheers anon

>>20117 Yea, so without using LoRAs, weights, breaks or anything, I was never able to generate a non-hires fix NaN error/Black Box image on fp16 So for fp32, do you want me to use special syntax in the prompts without LoRAs just to see if I can get one to happen or do you have any special instructions?

>>20118 btw I was >>20011 >>20014 so I know what the game plan is, just not sure if you need anything different for fp32

>>20118 Without special syntax is important if I want to compare against Naifu. With special syntax (besides 3rd party extensions) is fine but will likely take more work to debug (if NaN w/ fp32 even happens at all).

>>20120 Ok, I will at least do a repeat of what I did yesterday but with fp32. I'll get a NaN Black Box gen with weight/BREAK and I'll do a generate forever for 15 minutes with none of the special syntax. I'll get started in 10 minutes and have something for you within 30 minutes after

>scan, (traditional media:1.3), (sketch:1.2), (extremely detailed CG unity 8k wallpaper:1.2) >BREAK >hakurei reimu >Negative: (worst quality, low quality:1.2) >15 minutes of generate forever produced nothing adding the Zankuro and Ponsuke LoRAs and letting it run for another 15 minutes

>>20105 hi seraziel anon i just found the momiji from june and i never thought i could nut that hard

>got banned again for posting how to filename filter the stupid demon fag on /g/ Next time I get unbanned I'm gonna add it in catboxes.

>>20124 Demon fag? Do I want to know?

>>20122 >another 15 minutes >got nothing again Not sure what to do differently to try and force an error, maybe see what the other guy did to prompt the error

>>20125 some mentally ill retard that should be trip fagging but refuses to, I made a list of all possible ways to filter him and he reported me (because everyone else hates him too so who else) and got me banned

>>20123 Thanks! reminds me I still haven't finished it, especially because last time I baked with picrel like a tard (hmmm I wonder why I had to lower batch size, must've been pulling (clueless)) I have to finish the Inori lora (retagging all eyes and testing shit since it's basically multi-concept), finish that shishi juuroku lora properly, and then rebake everything else lmao I'd like to retrain seraziel (and berseepon09) on zSNR vpred furry models but that means making furry tags too all of that while not being a neet anymore here's some gens from that lora btw, two recent with other style loras on top and one older, bare NAI

>>20124 >getting banned for that reason alone Thjs demon fag dude knows a mod, calling it now, you can't be banned for that one silly reason.

I have a VPN also that works on cuckchan but I'm not gonna burn it on him yet

>>20128 rip, godspeed also i meant the momiji that seraziel made kek, sorry i didn't know you posted one

Sorry Catbox anon, I was unable to force a NaN error on fp32 with all the things I tried that were not extensions.

>>20130 fuck, couldn't help myself, probably should've used my hotspot to post the filename filter

>>20131 ah fuck, I didn't check what he posted for a while makes sense

>>20133 /g/ is absolutely not worth bothering with, that shithole is gone long ago. Don't burn your vpns for that shit, the working ones are getting ry scarce on cuckchan. Although I did get banned myself from /h/ for shitposting when that comfynigger started shilling his filth.

what if catbox with 200mb of text in metadata

>>20135 I saw your deleted posts I'm still shocked that only your stuff was nuked and not the shills. I'm not too happy with how this is turning out >>20136 for what purposes?

>>20137 I probably went too hard with shitposting so it wasnt much of a surprise kek.

>>20137 >for what purposes? fucking with jeets

>>20139 a combination of resolution size and spicy metadata should do the trick then, have it take up the whole screen when they open the catbox link and have a bunch of hanging troonjaks in the metadata

>>20134 tons of extremely high quality stuff including polt, new holo x myuri and more importantly momiji

kek this is the hardest I've seen Cumrag get blown the fuck out in such a short timespan in a thread

>>20142 >from cumfy to cumrag LMFAO

I'm trying to understand something. Whenever I prompt for gym-related clothes (with gym in the negatives) there's always a chance it will give me broken barbells, a gym background, etc. Why? Is it a dataset + tagging issue on NAI's part? Is it CLIP still being CLIP despite """hacking in""" booru-style tags? Both? What do you even call this kind of "contamination"?

>>20144 It's probably a dataset issue but I can try testing something on my finetune, give me your prompt

Made a style lora for hunyanyac/gnsisir https://mega.nz/folder/dfBQDDBJ#3RLMrU3gZmO6uj167o-YZg Full size grid included I have yet to find a checkpoint that can do his flat style justice, every one I've tried just gets mangled and based65 really really wants to add soft shading There will probably be another version if I can't get the tsurime eyes he draws to come out with this one.

>>20146 that looks really good!

>>20145 Not home right now but it's literally just gym uniform/shirt/shorts with gym in the negs. I find that upping the weights for the negs or being too specific (eg barbell, weights, weightlifting, etc) is counterproductive. Either that or SD is just very spiteful.

some lewd tests

>>20148 Yes no the very mention of gym in the tag and negging gym is gonna cause all sorts of problems. You will need to specify the type of clothing like “yoga pants” or something of the like

>>20150 Which is exactly why I want to know if this is a dataset issue, a CLIP issue or whatever the fuck. There are no other tags to refer to gym shorts, dolphin shorts aren't the same and the tag isn't as "strong". It's the spats vs bike shorts issue all over again. You can recreate gym shirts and gym uniforms but it's a waste of tokens imho.

>>20151 its probably leaning to a CLIP text encoder issue then

Oh Catbox anon, thought you'd might wanna know about this. I got a NaN'd error without any special syntax but on fp16. Now at first, I thought it was because of my finetune, but I decided to keep the prompt exactly as is and switch to nai-full and eventually I got a NaN error. I have it saved for when you get back on in case you need it.

>>20153 I'm here

>>20154 https://litter.catbox.moe/01r4n4.png (Euler a) https://litter.catbox.moe/n0rfk8.png (2M Karras) Thought I would include getting it in two different samplers

>>20153 i still have to figure out if catbox anon is the actual catbox owner or if he's the guy who made the catbox upload script

>/g/ got datamined for an AI detection application https://github.com/Urist-Mc-Urist/AI_detector lol >>20156 catbox upload script

>>20157 >/g/ faggot's hubris and pride will inevitably fuck us over when he ends up selling his shit to arthoes sites But can this faggot beat adverse cleaner?

>>20155 Okay, so I get NaNs in the webui, but now I realize the dtype env var for Naifu only applies to the diffusion model, not the VAE. They forced it to run in fp32. Changing it to fp16 is giving me a different error but I think I need to throw in some autocasting somewhere so there's still a bit more for me to look into. This means finding an image that produces NaNs with fp32 would be very much sought after though.

>>20159 >They forced it to run in fp32 Wait so they did DK64 the VAE?

>>20159 >This means finding an image that produces NaNs with fp32 would be very much sought after though. Tell me the prerequisites, I have a few hours and want to contribute for once. I don't have NAIFU anymore but I can go fetch it if needed. No special syntax, no loras, NAI VAE (duh), no weights, what else? Pre or post upscale?

>>20160 I need to figure out the current error I'm getting but if I do and if it throws NaNs I think it basically confirms they DK64'd it >>20161 pre upscale would be preferred but you could also just generate the image with an absurd resolution from the start and that should achieve the same thing

>>20162 The only alleged fp32 error we know was with that anon having LoRAs, im starting to think they did DK64 if Naifu has it force set to fp32. I've been running a generate forever for at 2.1secs a gen on a 4090 with the same prompt and I haven't gotten a black box yet in 250 images.

>>20162 >pre upscale would be preferred but you could also just generate the image with an absurd resolution from the start and that should achieve the same thing I lost my very first gens so I can't confirm if I was using NAI or not but I do remember getting black gens when trying to gen at 1024x back when I was testing my GPU's limits

>>20155 >>20163 So, I get the NaNs in Naifu too. They also intentionally casted the samples passed into the VAE to fp32 which was the reason for the runtime error I was getting. I swapped out the NAI VAE for one of SDs and it worked fine. Considering the rest of the pipeline in the VAE part just is taken directly from the original CompVis repo (like the webui used) I'd say it's pretty much confirmed at this point it requires fp32.

>>20165 Just to add to this I could go down the rabbit hole of seeing when exactly it throws the NaNs to see if it's consistent and attempt (a lot of emphasis on attempt) to modify the weights but my guess is it's not consistent and the VAE is just fucked

>>20165 >I'd say it's pretty much confirmed at this point it requires fp32. So it's unfuckable? What about the times when it blacks out even with --no-half-vae? Just a compatibility issue with LoRAs?

>>20165 Fucking rip, So how exactly are the VAEs made? Are they created from the finetune or are trained separately?

>>20167 My educated guess is the weights shifted by LoRAs are too much for the NAI VAE due to how unstable it is so it can't account for those changes >>20168 I believe separately. It's actually been around for years back when GANs were the thing. I also just now realized the kl-f8 anime VAE is finetuned off one of CompVis' VAEs, lol. https://github.com/CompVis/latent-diffusion#pretrained-autoencoding-models https://github.com/CompVis/stable-diffusion/tree/69ae4b35e0a0f6ee1af8bb9a5d0016ccb27e36dc/models/first_stage_models/kl-f8

>can’t check the furry diffusion discord on my iCuck while on the shitter because of AppStore policy on restricted servers >Don’t want to install DisCuck on my Graphene phone First World Problems are so fucking annoying

>>20169 If someone wants to train a VAE I think this is the place to start btw. It might not actually be that difficult. https://github.com/CompVis/latent-diffusion#training-autoencoder-models

>>20169 >>20171 Even when /hdg/ and /sdg/ were good... wait no, decent... wait no, fine... wait no.... even when they were not complete garbage - yeah that's better - NO ONE ever talked about or made a guide on how to make a VAE despite plenty of people asking, I wonder why

>>20172 Well one thing to note too is even Stability finetuned off the kl-f8 CompVis VAE: https://huggingface.co/stabilityai/sd-vae-ft-mse-original#decoder-finetuning Which originally was trained on this dataset: https://storage.googleapis.com/openimages/web/index.html I'm actually gonna look thru the NAI leak code again just to see if there's any indication how they trained/finetuned their VAE but I don't recall seeing anything last time

>>20173 You have both parts of the leak, right?

>>20174 yes but part 2 is just models, particularly the text ones and a bunch of the test SD ones, nothing really useful for our purposes

>>20169 >>20171 The reason I ask is because when I have a base model freshly trained on top of nai-full, I have a very vibrant yet sharp, and kind of harsh in some spots, color palate compared to NAI and wanted to see after I get the finetune to a more stable state, I wanted to train a VAE of the color palette (assuming the color improves the further I improve the training). >1st Image - old revision base finetune with NAI VAE >2nd Image - old revision base finetune with VAE Off >3rd Image - Current Revision Finetune + merged model with NAI VAE The old revision is pure ufotable screencaps where the current one has more stuff added in so I used the pure model one to demonstrate the colors but there is more visible training jank compare to the current one.

>>20176 oh and when I merge the fresh model with a checkpoint merge, it inherits the washed out color of not having a VAE on or leaving it on Automatic. Hence the lack of a 4th image with the VAE.

>>20162 >pre upscale would be preferred but you could also just generate the image with an absurd resolution from the start and that should achieve the same thing Started genning with a simple prompt at batch size 8 and I accidentally got a few bangers, no failed gen yet

>>20173 Couldn't find anything, there's a bunch of configs and commit history for the finetune but nothing specifically related to the VAE from what I can tell

>>20178 >>20162 Retarded question but the seed should be random, right? Also I should probably knock down the step count to 20

>>20180 I don't think it matters, i had mine on 30 steps, and yes do random seed, its not gonna spontaneously NaN a successful seed

>>20180 Yeah that's fine. I'd honestly say there isn't much worth in it now that I know the NaN issue happens in Naifu with fp16 but knock yourself out.

>>20179 I thought it was because I was a codelet but I felt that most of the stuff on that github leak part of the leak wasn't very useful for people wanting to reverse engineer the training process.

>>20182 >I know the NaN issue happens in Naifu with fp16 How did you get it to use fp16?

>>20184 It reads a DTYPE env var but that only applies to the diffusion model. You need to cast the VAE to half and the returned samples to half because they always expected the VAE to be in fp32. I don't know why the repo is on GitHub but it is, these are the two lines you'd need to change. https://github.com/Limitex/NAIFU/blob/1ca79a9e1a9d10e11af064c4a782f4ad208cd033/hydra_node/models.py#L231C70-L231C70 https://github.com/Limitex/NAIFU/blob/1ca79a9e1a9d10e11af064c4a782f4ad208cd033/hydra_node/models.py#L466C80-L466C80

>>20183 All the finetune code is there but to what extent that's useful I'm not sure. Penultimate CLIP and bucketing were the two major things and those were made public by them directly. I think a couple notable things they implemented were tag weights based on frequency (they literally comment it's to prevent reimu/miku from polluting the dataset) and resolution dropout (???)

>>20186 >they implemented were tag weights based on frequency That would be kind of useful for me since sometimes prompting 1girl blond hair gives me a saberface and if it gets annoying I have to waste neg tagging Saber. >resolution dropout the fuck is that? Infact, what is meant when referring to dropout, such as caption dropout?

I am going to do the forbidden and see if I can port some cumfy nodes to the webui https://github.com/city96/SD-Latent-Upscaler https://github.com/Extraltodeus/noise_latent_perlinpinpin

>>20188 >I am going to do the forbidden and see if I can port some cumfy nodes to the webui https://www.youtube.com/watch?v=tHXEXBsvTKg

>>20149 cute! catboxen?

>>20093 your cunny and cunny juice, sir

finally starting hydrus what the fuck is that "darkmode"

>>20151 if I prompt gym {uniform|shirt|shorts} and put inside gym, gym equipment, in the negs, then I don't get any broken barbells in any of the 16 gens I tried on based64v3

>>20194 what do you mean?

>>20196 Hadn't seen the style option. Also, is there a way to import files in a way to keep the folder structure I have? all of this is pretty overwhelming for someone who just wants to edit tags in a more streamlined manner

>>20197 While Hydrus does not keep your folder structure, you can group your imported image was special hydrus meta tags called group namespaces. Just highlight all your imported images and make a tag in the following format >group:name >ex: Series:'Name', Author:'name' Genre:'Type' Also, sorting by "Import Time" will at least respect the filename order of the images from the folder(s) you import.

>>20198 (Me) if it wasn't clear (which I don't think I did), you can make your own custom groupings so long as you follow that syntax so you could use that to group your stuff in folders that you had and make multiple subgroups if you need that level of autism orgganizing. You can pretty much do whatever you need.

>>20198 >>20199 Okay, thanks, I think I found that.

Just found that you can set "scheduler type" in A1111 master So if I set scheduler type, it seems to prioritize it over your sampler choice, so it will choose exponential even if you set SDE Karras as your sampler I think catboxanon said that exponential should be the best one for SDE, is that correct? And what is the "automatic" sampler for SDE?

>>20190 https://files.catbox.moe/19ffvm.png https://files.catbox.moe/pqwgth.png https://files.catbox.moe/7ke50e.png and here's some extras, from img2img https://files.catbox.moe/xlzv6o.png https://files.catbox.moe/p8th7z.png https://files.catbox.moe/v4z4un.png

yeah, exponential is best for dpm++ 2m sde. automatic uses whatever scheduler is the default for the current selection (most just use the vanilla scheduler)

>>20203 Meant for >>20201

>>20203 I mostly used SDE Karras, not 2m sde Interesting how distinct karras looks

>>20197 You can use regex during import to save your whole folder structure as namespaced tags automatically. In the tagging dialog during import, switch to advanced under your desired tag service and in the quick namespaces section click add. There add your namespace name to the first field, ie. folder and add (?<=^D:\\Folder\\Folder2\\).*(?=\\.*$) to the second field, where the D:\\Folder\\Folder2\\ part is the part of the path you want to exclude for the tag, so if your folder structure is like D:\Folder\Folder2\Cute\Whatever, then the tag will turn out as folder:cute\whatever. If you then want to search every file in the cute folder, you just search "folder:cute*". The simple tab also has checkboxes where you can just save the first or last 3 folder names as namespaced tags. Also if you don't want your folder tags clogging up your tag display on the left in gallery view, you can go to tags > manage tag display and search, where you can turn off specific tags and namespaces when you click on the tag filter for multiple file views button. You can still use the tags to search and they still pop up in autocomplete.

>>20206 there we go, wish I knew that

>>20206 Thanks. Also, is there any good way to do grouped tag work? For now I'm stuck using archive/inbox and favorites to hack up something, but I can't even seem to unfavorite multiple images at once so it's kind of tiring

>>20208 Ctrl + Click multiple individual images or Shift + Click an entire row or section and you can do tag multiple images at once. Also, in Options => Thumbnails, make sure ctrl-click and ctrl-shift focus thumbnails in the preview window checkboxes are clicked so you can see what you currently click when grouping together.

How does one think that it's fucking reasonable to adverise ui name in filename I hate cumrag so fucking much

>>20208 Yeah, just select multiple files and press F3 to edit tags for those files. You can double click on an existing tag in the editor and it will give you an option to add or delete that tag from every file in the selection, or just add your own. To mass edit favorites, select your files, then right click > manage > ratings.

>>20210 At least you can just filename filter them. So long as Cumrag gets visible push back and mock him by telling him to buy ad space every time he posts, that's all you can really do. And then make sure you push back the one or two retards trying to shill on /h/. They haven't shown up today so its probably fine for now.

>>20206 Going back to this, is there a similar system for export? I can only set things for the filename, not for the folder structure

>>20213 nevermind found it, it just didn't want to type "\" for some reason

>>20213 I don't think there is. If you plan on exporting regularly to the same folders, you could probably set up an automatic export (you can set it up to be manual too) for each folder in file > import and export folders. >>20214 Wait, it actually works if you add \ to the filename?

>>20215 >Wait, it actually works if you add \ to the filename? Yep, it's even on the help page only thing is that it makes everything lowercase for some reason

>>20203 Which sampler/scheduler would you personally use in general? I'm doing bunch of xys now and it feels like 2M SDE is kinda bad, although exponential does seem to cause the least amount of fuckups with it when I tested with high cfg. SDE Karras and 2M Karras look better at least.

>>20216 Damn, I just learned something new today lol. >only thing is that it makes everything lowercase for some reason Also noticed that if you use a tag for the filename and it has "\" in it (like the folder tags with the whole folder structure), it will get converted to "_". That's kind of a shame.

>>20217 >SDE Karras and 2M Karras look better at least. This is what I use as well. Hard to recommend stuff when not all them are really comparable (some intentionally by design work better at fewer steps and/or lower/higher CFG). What I've found really damn good though is using the Restart sampler in the hires fix pass. It's mediocre for the initial sampling but for hires it's great. I've meant to make some comparisons to post here but haven't gotten to it.

2M Karas for standard gens SDE Karras for Img2Img Upscale Euler A for inpainting

>>20219 >Hard to recommend stuff when not all them are really comparable (some intentionally by design work better at fewer steps and/or lower/higher CFG) Yeah I understand, that's why I asked your personal preference. >>20220 I used SDE Karras(genning/upscaling) and Euler A(for some of the upscales and inpainting) 99% of the time but those xyzs make me reconsider 2M Karras. I wish there was a way to make XYZs with SDE at 2x less steps compared to other samplers

>>20201 You could do this for months in ComfyUI.

>>20222 Hello cumrag, why haven't you bought any ad space yet?

>>20222 You could tongue my anus for months.

>>20222 Why do you even come here?

>>20225 4cucks isn't sucking his dick anymore and it's street shitting hour for the SAI pajeets.

>>20226 Yea but, they know that only a small amount of people come here still and we all hate cumrag

>>20227 Stockholm syndrome then.

>20222 wow, you could in the webui too! https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/10649

>>20226 >4cucks isn't sucking his dick anymore /g/? I thought place is a lost cause

>>20230 Yesterday he got blown the fuck up the moment he showed up https://boards.4channel.org/g/thread/95459181#p95459983 /sdg/ is still a lost cause because the avatarfags don't get necked so nothing useful gets talked there.

Im not posting gens on /g/ anymore now that people have begun scraping the threads. I normally wouldn't care but with the amount of trash getting attracted to that lightning rod of a place, I don't want to contribute to someone elses work.

>>20231 fucking kek, that's brutal. although there are still some of his simps there. Is that shit he posted actually any useful though? >>20232 I posted there once or twice a week but yeah it's trash. >now that people have begun scraping the threads dunno I always felt that it's a given, especially considering it's /g/. can't do shit about that

>>20233 >I always felt that it's a given I feel disgusted having my gens be in the same folder as cumrag and debo

>>20202 thanks! nice gens!

thanks to whoever delivered that follow up fake catbox lmao

>>20193 cute

>>20093 >>20191 Joining you, sirs

>>20238 Well I kneel, good job.

I got about 50 image sets left to crop… The auto crop I used to use for some reason isn’t centering the face for the clean square cut anymore so I have been stuck doing this shit manually Fucking kill me

>>20206 does hydrus have a script for scraping metadata on import yet? >>20220 for me, it's euler a for the base gen and ++2m karras or ++3m sde exponential for hires fix. euler a does a better job of smoothing over the malign noise that often crops up during lora mixes.

damn bitch it's just Barq's® Root Beer

Reading above on the posts I missed, I think some of you get way too angry about some things. This thread is like super slow still so you could ignore the couple posts that bother you instead of getting PTSD from /g/ .

>>20243 I only really get mad when they start showing up on /h/, I could careless if /g/ burns, the random dude that shows up here I have already been able to ignore him after a short reply questioning his efforts here.

>>20242 MUG Moment™️

>>20245 Has anyone done loras of the cunny rock?

>>20245 interesting take on a root beer cream float

>>20246 There's one on the vtai rentry https://rentry.org/vtaiLoRAs#holoadvent-koseki-bijou https://civitai.com/models/129972/koseki-bijou-or-hololive

>>20248 Is it good? Any examples? Can't really whip up any at the moment myself

>>20249 Just some initial tests, it seems alright but they pruned tags for character details which is dumb

>>20250 Does it do nsfw well? I've had issues where some loras forget the character when you try to get them to do nsfw Thanks for testing it by the way

>>20250 Unfortunately because of the shitty tagging the clothes are stubborn to come off. (nude, topless:1.2) did the trick

>>20252 Yeah I've had that problem before with other loras. Thanks for testing it out.

>>20250 >>20252 I like flat loli bijou more....

Rock hands, that's a new one

>>20255 why is her skin so shiny

>>20256 just the style lora she a little confused but she got the spirit

>>20254 >tfw primarily a cowtits hagfag but sometimes go on an all night cunny binge like im a werewolf or something. y-yeah I like both I guess.

>>20259 I primarly like vanilla sweet stuff but sometimes I'll binge on the most degenerate shit prompting for hours or days then not consume any porn for weeks like my body needs to recover from how fucked up my last fapping session was

the only thing that terrifies me is the crunchy cereal bar

>>20241 >does hydrus have a script for scraping metadata on import yet? https://github.com/space-nuko/sd-webui-utilities/blob/master/import_to_hydrus.py I just edited out the parts that create notes with a filename and replacing " " with "_", because I didn't like those.

>>20260 lol yeah scrolling through my gen folder shows exactly this play out repeatedly

>>19867 New try. In pink on tensorboard. Seems to have gone a bit better (a tad less melting, but still a little), no more orange eyes thanks to retagging. However pruning the outfit tags didn't work out well. May also reduce the weights on the anime side of dataset since it bleeds a bit of the style. Full grids: https://files.catbox.moe/hjav0n.png https://files.catbox.moe/fiomje.png

Did the A1111 webui favicon stop loading for anyone else? My tab feels naked without the shitty little gradio icon.

>>20265 Gradio broke it a long ass time ago and it was only fixed in 3.33.0. https://github.com/gradio-app/gradio/pull/4369 Latest stable webui versions uses 3.32.0. Upcoming release is probably going to be using 3.41.0, whenever that comes out.

>>20266 Thanks anon

jeets are going really fucking hard lately.

>>20268 luckily you're here to complain about it

I need a better system for keep tags, khoya has only has a meager global setting, and easy scripts manages to do it by folder, but still no dynamic one. Like telling it you want "X,Y,Z" tags to always be kept if they're here Currently on 1, if I move to 2 it'll fuck with images that only have 1 main tag

For the anon who was using multidiffusion upscaling and losing details sometimes, I've gotten a pretty good result with the following: - tile controlnet as usual - reference_only controlnet - adding this made the biggest difference to preserving style & detail - nearest upscaler instead of ultrasharp - avoids the smooshing effect of ERSGAN upscalers killing some texture details - using 0.6 denoise because of nearest upscaler to avoid blocky artifacts - using noise inversion because of 0.6 denoise so that it preserves the original https://imgsli.com/MTk5NjM0 It added too much texture to the background, but I don't mind fixing that in PS.

>>20271 >- nearest upscaler instead of ultrasharp - avoids the smooshing effect of ERSGAN upscalers killing some texture details I honestly still haven't tried multidiffusion but is nearest really better than something like remacri for example? I dislike ultrasharp myself but remacri doesn't seem to kill the details. 2M or SDE karras also should be better at keeping details compared to euler a, at least that was my experience with it when upscaling. maybe I'm wrong though, I'll try multidiffusion later myself

>>20272 A long while ago I tested remacri vs ultrasharp and consistently preferred ultrasharp, so gave up on remacri. But I'll give it another go for this kind of use case. I was going to say "Multidiffusion upscaling suck with any sampler other than Euler or Euler A." But I did a grid to test now with remacri and it actually looks pretty good with everything other than DDIM. Maybe remacri just works better? I'll have a look at some of my older upscales to see if they get messed up with 2M or SDE. I recall 2M Karras giving "too many" details when I first tried it. Yet another knob to tune for future gens.

Finally fucking finished cleaning up the heavens feel images I had slotted for the month, still got some image composition tag fixing that I have to do but if I finished all that this week, I will be focusing on just trying to make improvements to my finetune settings and getting a new model revision out.

Wish I could pay anonymously for one of these colab sites/alternatives I want to gen shit online but I seriously do NOT want any of these sites to have my name, my billing info or my cellphone

>>20275 Why do you wanna gen shit online instead of locally on your computer?

>>20276 Because my computer is a mid-tier laptop from 2019 with a 1650?

>>20277 Oh fair. For some reason the way I read it I thought you mean to gen on the go or something.

last time I checked a furry checkpoint the quality was shit. where is the new furry science happening?

>>20279 Not sure on the specifics honestly, just that there are two allegedly good finetunes that others have mixed with porn models to improve genitalia or some shit

>>20279 https://huggingface.co/lodestones/furryrock-model-safetensors/tree/main/fluffyrock-1088-megares-terminal-snr-vpred You need this extension plus adding the yaml to your model folder https://github.com/Seshelle/CFG_Rescale_webui Even just as a base model it's pretty usable but it's a furry model and prompting for humans doesn't fix that. If we had a danbooru equivalent I think it would easily beat out NAI. I haven't tried any fine tunes but it would be interesting to see how the different tagging affects how you prompt but it would suck if using e621 tags just made things more furry

>>20281 >said a stupid remark on /h/ about finetune that I probably shouldn’t have >get exposed >people now talking about me Do I have some fucking fan club or what?

>>20281 Didnt mean to tag you in the post above, but wanted to know if the hugging space has anything regarding training parameters

>>20282 >Do I have some fucking fan club or what? Welcome to the club I have a small but dedicated group of schizos that often accuse other anons of being me or sperg out about me while I watch from the sidelines

>post catbox for anon and a gen >leave my home and come back hours later, check 4cuck /hdg/ and that sperg replied to me again I guess this dude's obsessed with me for some reason.

>>20285 Is he just asking for catboxes or is this the fucking that trolls /hdg/ with cumrag shill

>>20286 >trolls /hdg/ with cumrag shill dunno he keeps telling me I need medication for some reason

>>20287 did you take your meds though

>>20288 yes, my protein shake.

>>20288 Idk about that anon but I just took my favorite medicine

>>20290 That sounds like a good idea right about now

>>20291 It's the GOOD stuff

>>20292 I just have ye ol reliable Johnnie Walker in the cupboard, though its still a bit too early for that right now

>>20293 my motto is 6 pm was yesterday

>>20280 >>20281 thanks. curious to check it out

>>20273 How many steps/cfg was this? DDIM got fucking destroyed lol

>>20264 Another bake. Green is last on tensorboard. It's better without the pruning. Orange eyes are kinda back for some reason, but now that it's tagged I can probably just put it in negs. Anime style doesn't bleed as much, but it still struggles for the WhiteVoidDress (but not for the version with cape lmao). I'll try running it for longer with a bit more restarts (for now I had 1 restart per epoch, I'll bring that down a little). I'll also probably switch to khoya_ss (I'm on bmaltais) so I can properly use the toml files for dataset, in order to use the mean number of keep tokens for each folder. I'll finally bump up the repeats on voidthreads and monster, same with whitevoiddress just in case. Grids : https://files.catbox.moe/5ugprh.png https://files.catbox.moe/tnorh5.png

>>20297 I've also been running with --zero_terminal_snr this whole time, not exactly sure if it's even applied

If I can't revive the game then I will revive the art style.

>>20299 I dont know what this is

>>20300 The most fun predatory f2p gook arena "sports" tps of my childhood.

>>20301 I still dont know

>>20302 S4 League. https://s4league.fandom.com/wiki/S4_League

>>20301 >>20303 >2007 >childhood I was in highschool lol

This brat...

Sometimes I wonder if the AI ignores my negatives once in a while just to fuck with me

>>20307 You're not alone People think I'm a schizo when I tell them SD starts to ignore tokens altogether even after restarting

>>20271 How do I get a good upscale without adding a third nipple or fucked up fingers

>>20308 I have noticed that sometimes introducing a negative tag will add it in anyway

>>20310 It's a 50/50 for me, it's bizarre

>Today I will be productive and do things!! >But I'm feeling kinda down so I'll prompt a bit to feel better >End up prompting the entire day and got literally nothing done this sucks AI is a terrible influence on me

>forgot I had another batch of images I needed to crop >proceed to drink instead and smoke instead Fuck

I wish there was an easier way to control where the character is looking instead of having them always face the "viewer" or the camera

>>20314 Same, I find that the "looking" tags other than maybe up/down (and at viewer of course) basically don't do shit, especially looking to the side/away. This is probably where the controlnet rig is useful but I feel like that and controlnet in general kinda suck the fun out of pure t2i

>>20314 >>20315 (me) I feel like this is yet another issue caused by NAI's dataset and most likely sloppy tagging. The more I learn about Anlatan/NAI the more I view them as outright incompetent. Then again, even pre-leak I didn't have much faith in a company founded by covert r*dditors on 4cucks then moved back to r*ddit.

>>20316 To be fair, booru tags are just not that descriptive when it comes to directional movement

>>20317 That's assuming stuff is tagged correctly to begin with and that's rarely the case when it comes to anime boorus. Furry boorus (so mainly just e621) are MUCH better when it comes to tagging (but it's nowhere near as godly as people want you to believe lol, trust me) Tags like "looking away" and especially "looking to the side" are functionally identical to "looking at viewer" right now with very few exceptions.

>>20318 Yea those two are fucking aids Also >there is looking up and looking down tags >but no looking left or looking right Im gonna fix that in my model

>>20319 >>but no looking left or looking right You could replace those with away/side... but alas...

>looking away danbooru page >”This tag is deprecated and can't be added to new posts.” Uwotm8???

>>20321 What the fuck?

https://danbooru.donmai.us/wiki_pages/looking_away https://danbooru.donmai.us/wiki_pages/facing_to_the_side

>>20321 >all those times I negative prompted "looking away" and "facing away" didnt do shit oh god

>>20323 This is clearly a Danbooru issue but what the fuck lmao

>>20323 i can't believe novelai did this

>>20323 Oh my fucking god. Too ambiguous? I get it but that's so fucking retarded for multiple reasons. Broader tags help while searching and you could always use tag implications. Not only that but the old posts are never gonna get updated anyway. >facing to the side SPOT THE FUCKING ISSUE

>>facing to the side >"A character facing to their side with closed or covered eyes." Was someone fucking high when they wrote this? >"The character should be facing to the "side" relative to their body, that is, turning their head without turning their body. If the entire body is facing to the "side" relative to the camera, use from side instead." >examples are not exactly accurate across the board urgh

>>20327 Does Stable Diffusion in general even have the capability to understand tag implications?

>>20329 By proxy, yes. Same way it'd work on a booru. Or... when pruning captions.

excuse me for mentioning this, but is there any way to try cumragui prompt weighting on a1111? I was wondering how it works and if there is any situational benefit to it but I really don't want to bother with noodles

>>20220 What makes euler a better for inpainting compared to just using the same sampler? I use 2m karras for everything.

>>20332 Nta but in my experience it requires less denoise to make perceivable changes, it looks kinda soft and blends well with rest of the picture and it produces pretty clean eyes when you inpaint faces. It's bad for inpainting hands

I fell off in march, I'm just getting back into AI, are the based mixes still the king for anime girls? also easyscripts anon if you still post here your a fucking saint. I'm really impressed with how far you've taken it since I last used it

>>20334 based64v3 remains the champ, yeah

>>20335 embarrassing

>>20335 But counterslop-v3 makes the best waifus.

>>20336 only 857 million parameters prease undastandu

>>20307 A1111 is personally fucking with you

>>20309 I was going to say you have to enable the tile controlnet, but the metadata says you do have it. You have put the CN ending step to be less than 1 though, keep tile controlnet on all the time. I also don't see "Tiled Diffusion" in your metadata, I think that is actually necessary, because it uses tiles of a size close to what SD was trained on, so it doesn't cause artifacting (when combined with tile CN). One last thing: try Euler A, I know it seems like it'll suck, but give it a shot. With Tile Diffusion, the "latent tile" size can be multiplied by 8 to get pixel size. So if you wanted 768px tiles, then use 96. That's what I prefer, although 128 (1024px) is more efficient (faster) at low denoise.

>>20341 Thanks for the tips, I've basically been using controlnet tile to guide the first steps up upscaling to minimize the weirdness of latent upscaling which is why I had the ending step less than 1. I found that "just resize (latent)" doesn't error out from low vram while "just resize" does, so I had just been doing that instead of proper tiling.

>>20316 We would be eating out of a dumpster if it was not for the NAI leak. >>20332 picrel is cropped details from an earlier image with a variety of samplers. Euler A just looks better, I don't know why. The skin starts looking like sandpaper with odd wrinkles and bumps and there is unpleasant texture getting added to the plain color of the sheet. My uneducated guess is that the other solvers are "better", "more accurate" and end up trying too hard, hallucinating a bunch of detail that shouldn't be there. Euler A's lack of competence is its strength when it comes to upscaling. >>20342 With hands, I try fix them with gimp + inpainting on the low res image before upscaling. It's also worth doing an inpaint pass on the face before upscaling, gives a higher quality result usually.

>>20343 >do an inpaint pass on face before upscaling I always had a shit result when doing this, but I assume you do another pass later after hires/upscale?

>finally unbanned from cuckchan >don't even feel like posting on /g/ anymore despite having ban evaded the entire time weird but now I can do more useful things

>>20281 I'm plenty fucked up but thank god I'm not a furry. My gens look fried, doing something wrong. The furry discord gens look high quality, but not better than these threads. CFG rescale and vpred actually fixes darkness, picrel.

>>20346 use cfg rescale at 0.7 with trailing enabled do you have the yaml next to the model, with the same name?

>>20344 No extra pass on the face after upscale. Usually try tweaking the settings and rerunning the upscale, then combine my favorite two with gimp. first pic is original gen, second is face inpaint + arm inpaint

>>20345 Yeah same I think some of their banlist got fucked when the site went down last time, I was gonna post my stuff on /b/ but fuck that board in general.

>>20349 yeah it's a hellhole, it's either spammed by one of the local schizo avatarfag, or the baranigger kills the thread instantly (mods don't do anything and when questioned on IRC say "they didn't know" lmao)

>>20350 Yea I'm just gonna go back to my not so tinfoil conspiracy that the mods are looking for an excuse to kill all the /ai/ generals and letting all the schizos run amok is one way of going about it, cuckchan won't have an /ai/ board at this rate.

>>20345 /g/ is unironically tolerable at euro hours

>>20347 Got this in models: fluffyrock-576-704-832-960-1088-lion-low-lr-e125-terminal-snr-vpred-e98.safetensors fluffyrock-576-704-832-960-1088-lion-low-lr-e125-terminal-snr-vpred-e98.yaml Maybe I chose wrong model. Using 0.7 CFG rescale, but the trailing checkbox is only for DDIM?

>>20346 Dunno, seems fine for me. It's fun trying artist tags

>>20346 what am I even looking at?

I consistently get black box failed inpaint with NAI vae right now. Seed not fixed. Started happening after I disabled controlnet and now works fine after I enabled controlnet back. Pretty weird. I had a similar problem when using adetailer with controlnet, it started black boxing face inpaints like that. Yeah I know nai vae isn't recommended for inpainting but still might be an interesting oddity. --no-half-vae --disable-nan-check are enabled

>>20353 >but the trailing checkbox is only for DDIM no, it's just a way of correcting an off-by-one error on the noise scheduling. It doesn't have much impact, most of it is from the cfg rescale itself when the model is loaded, check in cmd if it says "running in v-prediction" or something similar

>>20355 It's just showing that the vpred model can generate pure blackness, while based64 can't. SD was trained wrong and tries to always have a medium average brightness across the generated image. That's why you can't generate night scenes without it putting bright lights somewhere.

>>20356 I get problems with NAI.vae when img2img and inpainting as well but I dont have fp32 enabled Whats a good vae for inpainting?

>>20358 is fluffyrock custom base model or just the finetune model?

>>20357 Ok, will tick box thanks. I got "Running in v-prediction mode" Creating model from config: models\Stable-diffusion\fluffyrock-576-704-832-960-1088-lion-low-lr-e125-terminal-snr-vpred-e98.yaml LatentDiffusion: Running in v-prediction mode DiffusionWrapper has 859.52 M params. Loading VAE weights specified in settings: models\VAE\vae-ft-mse-840000-ema-pruned.ckpt >>20354 Ok, I can repro those. Looking at some gens on civitai it seems that including artist names is the way to get decent gens. Still doesn't make me feel sad to use Based64v3 as my main model. The furry models have gotten a lot better than when I first looked at them though.

>>20356 This image will fool /a/ 1000%

>>20356 It's especially bad for inpainting since it's the one that causes desaturation on each pass. >>20359 Use one of the SD or kl-f8 ones

>>20349 i miss the spirit of the old /b/ threads >>20362 nah, a human artist would never mix up madoka's outfits or glaze over her soul gem like that

>>20359 >but I dont have fp32 enabled Well it black boxed for me even on simple txt2img when I didn't have no-half-vae enabled. >>20362 Awful hand holding would give it away for sure but I was trying to fix it. Some errors with Madoka outfit too. Was mostly testing regional prompter latent mode and it seems to be working better than I expected >>20363 >causes desaturation on each pass Do you mean for the inpainted area?

>>20360 Fluffyrock is a base model, it's not a finetune of anything. Fluffusion is one too. but there are a bajillion versions of fluffyrock so it's a bit hard to navigate >>20361 seems good don't forget to use furry artist tags for quality (or the lora someone is making, if it released), and use e621 tagging vpred tagging is quite different, mostly in redundancy and weights

>>20364 at first glance they sure will, then the artcel seethe will commence >>20365 >fix image >wait for another Rebellion sequel update so it gets talked about on /a/ >pass the piece off as key art in the OP for a thread discussion

>>20366 Ok, prompting was the issue (and maybe the checkbox a little lol). Getting some decent gens. Dialing the furriness down, can get some cute anthro characters with cat ears, tail and colored skin.

Can we merge the fluffyrock model or is the vpred gonna be an issue?

>>20369 you can using the non-vpred checkpoint: futanari factor was done that way using the vpred checkpoint: this abomination was made : https://civitai.com/models/124655 (fuck there's a v3.0 gotta test it out)

could Based64 get some vpred to make night generations possible without lowra hackery?

>>20371 We'd have to a model that has vpred into the based mix

is vpred only a 2.0 thing or can you finetune a model with vpred on?

>>20371 >>20372 using supermerger's trainDifference, you could, see the second model in >>20370 >>20373 I don't think vpred requires architecture changes, so you probably could finetune a model using v-loss and zero terminal SNR

>>20374 If I could get some documentation or settings that I need to fuck with in Kohya (I’m still using the powershell scripts not the gui) I can try training a vpred version of my finetune to use in mixing up a variant of based64 from scratch this week

>>20362 why does /a/ get so triggered with any ai?

>>20376 its an /ic/ satellite with artcels and drawnigger

>>20343 >We would be eating out of a dumpster if it was not for the NAI leak. Haha yeah, for sure! Lmao no, fuck off. Maybe the lack of a base would've helped WD get farther than absolutely buttfuck nowhere due to infighting.

>>20378 Are you lost? This a new troll angle?

>/hdg/ is getting trolled now yup, I thought this guy was fishy, hope he likes getting mass reported

>>20381 Tongue my anus. I'm thankful for the leak, I just hate how sloppy base NAI is and how many issues could've been avoided had they actually put some effort into it instead. It makes you wonder what could've been if WD was "forced" to at least be not shit out of necessity.

>>20382 You are only saying that in current moment because you looked behind the curtain, and you should be glad they did more than Stability had RunwayML do with Laion5B. The other problem with your line of thinking is that WD had access to the leak, they had devs that could've read up what NAI did to make their own improvements, and they didn't capitalize on it. It's kind of pathetic when ControlnetDev knows more about NAI's training process than the WD devs, one who according to what people have said now works for Midjourney.

>>20383 >You are only saying that in current moment because you looked behind the curtain, and you should be glad they did more than Stability had RunwayML do with Laion5B. Like I've said, tongue my anus. Go shill for your not-so-covert-anymore-r*ddit company elsewhere, they couldn't even get their VAE to work properly. >WD had access to the leak WD existed before the leak.

you are a troll, i'm ignoring you, go back and shit up /g/ or something since you have resorted to dodging the central point

I don't know why the fuck I can't post to vtai so this is going here

>>20386 > I can't post to vtai what do you mean? I posted there not long ago.

>>20387 I get the post successful message then nothing. Or I get redirected to https://boards.4channel.org/vt/thread/0 after a few seconds. I tried on other browsers and other devices on my network but same thing. I would try on data but I'm banned for ban evasion apparently

>>20388 I noticed earlier some of my posts on /h/ got eaten up and disappeared into the void. Must be some kind of cuckchan issue.

test

spent an hour and a half trying to figure out why multidiffusion was taking eighty minutes to run a single midsize picture and spitting out garbage forgot i left the sd upscale script on

>>20375 I don't know shit but the other model anon linked had this in the description if you want to try merge b64 with the fluffyrock vpred model and do something similar: >This version uses polyfur-lion-e76-terminal-snr-vpred-e25.safetensors and fluffyrock-576-704-832-960-1088-lion-low-lr-e126-terminal-snr-vpred-e99 mixed with 526Mix to provide the model with even better knowledge and aesthetics. Polyfur is special, as it attempts to include both high quality photographs autocaptioned by MiniGPT4 in its dataset alongside furry images, so that you get the best of both worlds. The model is early in its lifespan, but contains enough specialised knowledge that it significantly improves the quality of this mix. >Unofficial merge of fluffyrock-576-704-832-960-1088-lion-low-lr-e126-terminal-snr-vpred-e99 and https://civitai.com/models/15022/526mix-v14 using train difference with the supermerger extension in automatic1111. >Train difference is explained here https://github.com/hako-mikan/sd-webui-supermerger/blob/main/calcmode_en.md#train, and is what allows this model to exist. TLDR, this model uses the a1111 supermerger extension to merge 526Mix with FluffyRock vpred using these settings: 526Mix is set as model A, fluffyrock vpred as model B, v1-5-pruned (sd base model) as model C, add difference as the merge mode, train difference as the calculation mode, alpha is set to 1, and the output is saved as fp16 safetensors. Is the mix for Based64-v3 a secret? Can't remember if it was mentioned in a previous thread.

>>20391 It hangs for me sometimes even in the best case. I think multidiffusion has a memory leak. Either that or ControlNet. Because sometimes it gets slower and slower then eventually sometimes CUDAs but if I kill webui and start again, and send the exact same shit through, it generates fine.

>>20392 IIRC, Based64 is supposed to be AOM2_Hard + (HLL3.1 - NAI_Full) AD@1, then previous output + Defmix Red WS@0.5 The later ones the author said he was practically doing schizo merge experiments with MBW so I don't think he ever disclosed it if he could remember, we'd have to check old threads

>>20394 Thanks, saving that recipe. Sucks that 8chan has no search ffs. Based64 (esp V3) turned out so damn good considering it didn't even do any block weight merge BS. I guess the exact mix doesn't matter though. Sounds like you take Based64 as model A, fluffyrock as model B, v1.5 as model C and then train difference with the supermerger extension. I might try that when I get some time, but I think my gpu won't be strong enough.

>>20395 >considering it didn't even do any block weight merge the sum of the parts yea, but AOM2 was block merged so there is that Wished he was still here, he kind of disappeared along with everyone else

>>20396 >Wished he was still here felt like he got disappointed that his newer merges flopped honestly though, I don't get the praise b64 gets from /h/. it's good but there were other reasonable aom2 merges out there as well that weren't overfit to hell. and the overfit pretty by default models got their uses as well unless you really enjoy stacking bunch of loras for your every gen

>>20397 Those lightning in a bottle models like Based64 and Anyv3 always seem to outshine anything else ever made by their creators after

>>20398 yeah that seems like it sucks good thing it'll never happen to me haha

I was on and off this thing for some time but what actually lead to the consensus that based64 is undoubtedly best one for hentai and there are no better models? Its lora compatibility and good nsfw? Has anyone actually done xyz plots and compare this shit?

where's aom4

>>20400 >Has anyone actually done xyz plots and compare this shit? nope, we're all counting on you

>>20400 it just so happened to have very good lora and embed compatibility that has yet to be replicated since >>20401 Never, because like the examples above, AOM3 fell flat on its face compared to AOM2

>>20403 >it just so happened to have very good lora and embed compatibility that has yet to be replicated since wouldn't NAI and aom2 have better lora compatibility technically? I don't think it's the main argument >>20402 yeah I expected that answer but I simply wondered how people came to this conclusion

>>20404 just a lot of gening tbdesu, i don't really bother with 1:1 testing anymore because it's a pain to do it at a meaningfully large volume

How do I set the aspect ratio buckets for 768x768 training resolution or is it done automatically when I set the training resolution?

Hydrus chads, is there a way I can filter a namespace tag to become a regular tag? Example would be turn "artist:Greg Rutkowski" "artist:Hirohiko Araki" to just "Greg Rutkowski" "Hirohiko Araki"?

>>20407 and to be specific, I mean when doing an export so I can train

>>20407 honestly it's the kind of thing I don't even bother learning and just use a python script for but maybe you can just filter for that tag, select all, put in the new tag, remove the old tag, and voila

>>20407 You can process your tags when exporting. >click on the "no string processing" button in pic rel >click add > string converter >click add >select regex substitution >regex pattern: "^artist:" (without quotes) >regex replacement: nothing This will convert all artist namespaces to regular tags. Alternatively, you could use notepad++ and do a multifile find/replace after export.

Omega good job

>>20411 Fucking image failed to upload

Ez scripts anon when you catch this. >update sdxl branch >worked before, doesn't work after update >main.py line 13 / 30 >config = json.load (f) >json_init.py line 293 / 346 >decoder.py line 337 / 355 >json.decoder.JSONDecodError: Expecting value: 1line 1 column 1 (char 0) 2nd part of the monologue is I loaded up the old version that you abandoned, worked fine, and it seems to be way faster in using the same settings for the LoCON's I was making in the new version, like 2-3x faster. I don't know if it's not using the alog right since I haven't tested this batch yet but was something I immediately noticed.

>>20410 >>select regex substitution >regex pattern: "^artist:" (without quotes) >regex replacement: nothing This is what I was looking for! Thanks!

>someone made an /sdg/ in /ic/ fucking kek

>>20415 gigabased dramafarmers

>>20416 it looks like it was cumrags that did it because they all flooded in and the thread got deleted

>>20398 >>20397 To be fair, it's better to have a lightning in a bottle hit than not have it at all. Companies and individuals spend decades chasing for that one viral hit or their lightning in a bottle success but never find it so it's better to have that achievement under your hat and be proud of it over having 10 flops and 0 successes.

>>20418 I'm not disagreeing, but some people can't handle only having that one hit and not be able to replicate another success after and just feel like it was just a fluke

>>20419 I call that the Ken Levine syndrome.

>>20415 That's hilarious goddamn

>>20420 To clarify, it's System Shock 2 vs the bi*shocks.

>>20422 >strike lightning in a bottle >twice >ruin it by making Bioshock Infinite a fate worse than death

Random ufotable schizophrenic lighting no new model yet, just doing tagging clean up by doing random gens to see what tags I need to fix

>>20423 >strike lighting in a bottle >ruin it by making 3 bioshocks ftfy

>>20425 True but Infinite was hated by even the normies that got sucked in by the original. That game killed his studio.

>>20426 Was it actually hated? I swear you can't look up anything about that game without running into porn or porn memes SFM was a such a huge fucking mistake.

>>20427 People don't like Overwatch but like Overwatch porn. That steam release fiasco made it very clear to everyone if you lived under a rock. >SFM Its the people that it attracted that was a mistake

>>20428 >That steam release fiasco made it very clear to everyone if you lived under a rock. I was in high school and heard my classmates talk about the game a few times but I didn't like 1 and 2 so I never even bothered to look it up or look up the drama surrounding it. I was too busy playing DotA all day.

I didn't prompt this

>>20429 >Dota over OW based

>>20431 I eventually bought OW because the same classmates promised they'd play it with me while I was stuck in the hospital for a good year and half and thought "whatever it's a shit game but at least I won't be lonely" I booted the game up a total of maybe three times and I played with them about... never. Great friends, huh?

>>20432 Feels bad man, had the same shit happen to me with Destiny 2 at launch. Got bullied into getting it after I quit D1 for being nogame the game, got met with the same nogame and my friends didn't want to raid or do try hard PvP so I quit again not long after.

>>20433 I almost bought Destiny 2 but didn't because I had nobody to play it I got gifted XIV by a friend because I really wanted a new MMO to play but I don't stay subbed because I don't have anyone to play it with and ERP isn't my thing

>>20433 I have 50 hours in D2 and I remember very little of what happened during all those hours. Gonna try it again as soon as I build a new rig. It's basically the borderlands-esque game I crave from what little I can remember but my 3570K hated it.

Civitai is cursed There really should be an anime tag so im not flashbanged with 3DPD gay bara sex or furry shit In desperate need of eye bleach right now

>>20436 I think your problem was visiting that shitty place to begin with

>>20436 There are some decent loras on there but its a fucking needle in a haystack deal, but yeah holy shit scroll down to get eyes assulted by furry anuses and gay dicks is never fun, I just blocked most of the gay furry bara tags and creators.

>>20438 Is it possible to share blacklists? Can you share yours?

>>20439 I don't see an option here (unless an anon can tell me where I can find one) either way here is my list, those blocked users primarily share gay and furry works, any straight 3DPD stuff isn't blocked (mainly cause I don't give a shit).

>>20440 I never had this problem, but it could be because I search only checkpoints

>>20436 The worst curse is getting a lora of a style/chara you like and looks decent in previews but realizing that its fried unusable garbage. I learned to mentally block furryshit a long time ago so I don't even see it kek.

>>20442 >>20442 This is why I never check that place for anything unless its the only place to procure something like a model or some shit. Speaking of which, what are some tried and true artist and character LoRAs that you've ever used and/or made by anons?

Worst creator that regularly gets 5 stars on Civitai? I'll start: Malcolmrey. Just awful. I've tried several of his LORAs/Lycos and they're all just absolute dogshit, that only work in the narrowest of prompts and checkpoints.

>>20444 >I've tried several of his LORAs/Lycos and they're all just absolute dogshit, that only work in the narrowest of prompts and checkpoints I thought that's how all non-NAI trained loras worked lol.

>>20445 I've trained my own on models that aren't NAI and they come out fine. I've also found some decent stuff on Civitai, but it's not super common. I'm too lazy to make my own character LORAs. I just wanted to see Megyn Kelly getting her back blown out, but w/e

Is anyone taking style requests? I was thinking Waterkuma might be a good one.

>>20446 Every realism lora I've tried was absolute garbage that worked exactly how you described before. Even the fried 0.6 shit for anime was much easier to use compared to that garbage. Even considering all its faults that get mentioned itt, NAI was a blessing imo.

>>20435 You're not missing much, anon. It was OK at launch, but like all the battle pass games, it's now garbage revolving around FOMO and with mechanics and reward systems that dictate you treat it like a second job. Just watch Morbius instead.

>>20443 >Speaking of which, what are some tried and true artist and character LoRAs that you've ever used and/or made by anons? Original CSRB, new takorin from here, fizrot also from here, seraziel from here I wish the guy who made these would come back and retrain the pre-LoCON ones, they're alright but they could definitely be better. https://mega.nz/folder/OoYWzR6L#psN69wnC2ljJ9OQS2FDHoQ/folder/D1hGQCZI

>>20449 >it's now garbage revolving around FOMO I'm only weak to Arknights' FOMO so I'm good

>>20394 >Defmix Red Seems like the author got a little meltie few months ago. Anyone got the model saved? Can't seem to find it.

kohya uses comfy now, he even made a node for his latest thing https://github.com/kohya-ss/ControlNet-LLLite-ComfyUI good times ahead

>>20452 I downloaded all of them, I’ll reup it

This took some doing but it was worth it. All my character loras updated https://rentry.org/zp7g6 >>20447 I'll see what I can do

>>20452 >>20454 https://pixeldrain.com/u/b1tp6mnr This is only Def Red, but I have the rest that I can try uploading another time

>>20441 Any good checkpoints?

>>20457 nope, I only end up downloading whatever /g/ and /hdg/ say they use if the only option is to download from civitai

>>20458 Damn I downloaded some loli japanese model that I like so i'm on the lookout for other models all the time

>>20453 More like comfy times ahead amirite Cumfy fucking sucks I hated using it and the only reason I'd ever use it is if A111 is not an option at all

>found a 'colab for complete retards' guide >Click it >Has more options, buttons and text than the lastben and nocrypt collabs this is very funny and amusing

Is there a guide on composable LORA or mixing LORAs together if I want to use three character loras to generate three character in one picture?

>this is posted >suddenly /g/ got really quiet >no back talking >no samefagging >no pajeet quality posts is it finally over?

>>20462 I remember someone made a webm where he showed off prompting multiple generic girls and then inpainted each individual girl with a different character LoRA Mixing LoRAs to my knowledge would not help you generate multiple characters in one scene. Regional prompting might be another way to accomplish that.

>>20463 snitched about what?

>>20464 How do I use regional prompting for that?

>>20453 Why should we care about something so insignificant?

>>20465 Honestly it can be so many reasons at this point But I think the shit he said last night might have actually made their way to the yes of someone who didn't like seeing him mouth off

>>20462 Just use regional prompter in latent mode https://files.catbox.moe/2ig495.png

I see this in master when I pulled 254be4ee Add extra noise callback But how do we use it?

>>20468 >But I think the shit he said last night Out of the loop, what happened?

>>20471 said he would kill auto1111 (the UI), but can be framed as wanting to kill auto1111 (Voldy) as that is his git username

>>20447 Here's v1 https://mega.nz/folder/2FlygZLC#ZsBLBv4Py3zLWHOejvF2EA/folder/rUkHjK6C Not great at the moment, probably have to do some pruning

>>20472 >>20471 >Could work together >Could use the relationship as friendly rivalry >Obsessed with shilling his UI everywhere and "killing" the competition sometimes I wonder if humanity is just doomed as a species

>>20472 >>20474 lmfao cumrag sounds like a petty jew (redundant) and acts like one too so

>>20474 Yeah I was trying to subtly talk sense into him because this retarded console war is just tiring and there could be a chance to somehow repair /sdg/. But yeah he simply decided to double down. Hope he got in trouble for this shit.

>>20476 The way he doubled down just felt like watching an angry child in the playground as an elementary schooler and being hell bent on breaking everyones’ toys. Although I will admit that whenever I see his avatar, I don’t even read his post, I just tell him to fuck off. Only when I see other people greentext some particular Interest take do I then go back and read.

>>20477 >The way he doubled down just felt like watching an angry child in the playground as an elementary schooler and being hell bent on breaking everyones’ toys. It really seems personal. Could he be so narcissistic that this is all due to a rejected pull request? Possible but it's so fucking bizarre.

>>20478 I mean, didn't vlad basically do the same? he was a notorious shitter in the a1111 repo before doing his "own" UI

>>20477 >>20478 I mean it's cuckchan People hold grudges for fucking forever over the pettiest shit so imagine if you took one of these autists that hold petty grudges and gave them savant-tier coding knowledge then that's ComfyUI

>>20480 >savant-tier coding knowledge Cumrag is affiliated with SAI, what are the chances that SAI does the coding and lets the sperg advertise it? Your company looks better in the public FOSSfag space if you "support" a project like that.

>>20480 I've had a group of autists that were absolutely mindbroken from when I used to shitpost in a /vg/ general and now they're out there desperately trying to seek me out in any post that reminds them on me like ack thinking everyone that ever disagrees with him is *kemi The True 4chan Experience :tm:

>>20482 Forgot to mention this was two whole years ago and they're still at it. For some reason.

>>20479 Yes but he shat the bed on a fork, but Comfy “made” a new UI. Emboldens him to act like a petty cunt at every turn even when dealing with valid criticism. >>20481 No he actually made it, he documented his shit on /g/ and there are screenshots and Desuarchive links.

>>20482 FGO or Fire Emblem?

>>20485 Hoyoverse is all I'll say

>>20486 Did you bait people into pulling Qiqi?

>>20487 No but I got baited into rerolling her when I started...

>>20488 lmfao

>>20486 >genshit I forgot ufotable is attached to make an anime project with them… fuck

>>20490 I would post a smug Hu Tao but I haven't played my wife ever since I pulled Ganyu. ... I haven't had a reason to play anyone after pulling Ganyu now that I think about it.

>>20480 >savant-tier coding knowledge is it?

>>20491 Genshin is actually pretty fun and I think the game has come a long way since launch now that it's in 4.0 but it has also attracted an insanely autistic fanbase AND haterbase both in EN and CN so good discussion about it is rare