Discussion of free and open source text-to-image modelsPrevious /ldg/ bread : >>101749563>Beginner UIEasyDiffusion: https://easydiffusion.github.ioFooocus: https://github.com/lllyasviel/fooocusMetastable: https://metastable.studio>Advanced UIAutomatic1111: https://github.com/automatic1111/stable-diffusion-webuiComfyUI: https://github.com/comfyanonymous/ComfyUIInvokeAI: https://github.com/invoke-ai/InvokeAISD.Next: https://github.com/vladmandic/automaticSwarmUI: https://github.com/mcmonkeyprojects/SwarmUI >Use a VAE if your images look washed outhttps://rentry.org/sdvae>Model Rankinghttps://imgsys.org/rankings>Models, LoRAs & traininghttps://civitai.comhttps://huggingface.cohttps://aitracker.arthttps://github.com/Nerogar/OneTrainerhttps://github.com/derrian-distro/LoRA_Easy_Training_Scripts>Fluxhttps://huggingface.co/spaces/black-forest-labs/FLUX.1-schnellhttps://comfyanonymous.github.io/ComfyUI_examples/flux>Pixart Sigma & Hunyuan DIThttps://huggingface.co/spaces/PixArt-alpha/PixArt-Sigmahttps://huggingface.co/spaces/Tencent-Hunyuan/HunyuanDiThttps://huggingface.co/comfyanonymous/hunyuan_dit_comfyuiNodes: https://github.com/city96/ComfyUI_ExtraModels>Kolorshttps://gokaygokay-kolors.hf.spaceNodes: https://github.com/kijai/ComfyUI-KwaiKolorsWrapper>AuraFlowhttps://fal.ai/models/fal-ai/aura-flowhttps://huggingface.co/fal/AuraFlows>Index of guides and other toolshttps://rentry.org/sdg-linkhttps://rentry.org/rentrysd>GPU performancehttps://vladmandic.github.io/sd-extension-system-info/pages/benchmark.htmlhttps://docs.getgrist.com/3mjouqRSdkBY/sdperformance>Try online without registrationtxt2img: https://www.mage.spaceimg2img: https://huggingface.co/spaces/huggingface/diffuse-the-restsd3: https://huggingface.co/spaces/stabilityai/stable-diffusion-3-medium>Related boards>>>/h/hdg>>>/e/edg>>>/d/ddg>>>/b/degen>>>/vt/vtai>>>/aco/sdg>>>/trash/sdg
How good are AMD GPUs for AI image generation these days? Has support improved much since 2022?
Blessed thread of frenship
>>101753064I don't own one, but reports are "its working" .. if you got the $$$ go NVidia always still.
>>101753017>three of my images in OPlol
>>101753207welcome to the club
>>101753064You /can/ do it, but its outright awful in support compared to Nvidia
>>101753207lemme guess, top right, bottom right, bottom left
>>101753064still worse than nvidia, all of the stuff is always made for nvidia and for AMD its always a "maybe", you get a lot less mileage for the VRAM because the optimizations are also mostly nvidia only things, it works but it will never be a good experience>>101753221yeah, not hard to guess, all the gemstones kek
>>101747335Prompt?
Remember that I'm always right and everyone that dare to disagree with me is naturally a schizo
>>101753346This but unironically.
>>101753373
my new house
>>101753017:^) flux has a baked-in tit remover.
>>101753465its not a remover, nipples are just not in the data set, probably some AI detected and put red censor blotches on em
>>101753449Show your steps :^)(what are the gens like from each prior step? the way all image generators right now work is they "paint" changes to the image, and each step lets you see the current result)
my new pet>>101753484my goodness, I would need to make 24 gens at 1-24 steps each for that .. no thanks
Why don't these companies just slap a shitton of memory on a pci-e card and sell it? Skip the whole graphics component. That 4060 TI 16GB is a waste. That 7600XT 16GB is a waste.
>>101753483We'll know soon enough.Basic working fact:All of this type of imager does a loop over and over, editing the same image.Each step is a view OF THE SAME IMAGE AS ALTERED BY THE NEXT STEP.The setup is a little weird, it has to be cut in resolution by a lot, because the system can't "see" a large area at a time.But get it in your head now. Steps peek at the canvas, THE SAME CANVAS. more steps, it's NOT A NEW CANVAS, IT'S BEEN "PAINTED ON".
>>101753522Aren't you curious to see at which steps which parts of the image were painted?
>>101753625You don't understand how extremely niche it's to run local AI models. It's just not profitable for Nvidia.Even among my 20 Computer Sci friends, I'm the only who ever installed SD.
bros I there any chance to get an Azure Dall-E 3 proxy to generate nsfw? It even blocks if I try tame shit as "bending backwards to camera" wtf. Any alternatives maybe?
>>101753625greed, what stability tried to do with the amount of parameters available to local, is the same thing nvidia did with VRAM. the difference is that one is a software solution, while the other require expensive as fuck hardware factories to produce. so breaking the cartel is harder.
OK so someone mentioned "attention"Can we hijack attention and use SD's attention, but Flux otherwise?
my new neighbourhood>>101753662no, since I know that roughly already... I seen it in the preview window
>>101753664I mean it's not only compsci nerds doing this? I am a retarded neet dropout and I installed sd. Plenty of hobbyists install sd and have decent gpus. It's not hard.It's LLMs that have extremely restrictive hardware requirements
>>101753665YESSSSSokay, what you need to do is overwhelm the tasksSo like get thisThere is a baked in-100 titsso, if it finds tits, it goes and reduces them, okay?BUTIf you can find a way to make it really busy, you can always get your tits back out.An example of doing this is art styles, which Flux basically blocks rn, but maybe we can break that.See, with an art style, it has to work a lot all over the picture, it has to paint most of it.So again, give it looooots of heavy work to do, and it will wind up with -100 tits being unimportant enough it doesn't get to it in any steps.
Do you think we'll always be at the mercy of Nvdia or will something change in our lifetime?
>>101753704Why doesn't the preview window save the whole image?(idk why, but my dev fp8 for comfyui doesn't have previews)
So there's no real difference between the fp16 and fp8 dev model, right? The clip model actually matters though. I have a 3090 so I can use the full models but I'm trying to get faster inference speed.
>>101753764I suspect that we will gain a better understanding of llm and image thingies, and that we won't need such beefy resources to do lots of stuff we want to do.
>>101753780>So there's no real difference between the fp16 and fp8 dev model, right?not that much yeahhttps://imgsli.com/Mjg0MDEy
>>101753709>I mean it's not only compsci nerds doing thisIt's not. But I guess it makes it more probable that you:>are on the spectrum>gen waifus because you're a virgin nerd>have the minimum technical know-how to setup SD (Gen z and Alpha don't know how to use computers because they're phoneonlyfags)>find entertainment on fiddling with numbers and sliders all day long>engage with autistic people in chinese cartoon boards to learn more>keep update with new AI techs
>>101753772>Why doesn't the preview window save the whole image?it just doesnt.. maybe there is a node that does you could make some cute gen animations of them (maybe an anon knows? I don't)>>101753772>(idk why, but my dev fp8 for comfyui doesn't have previews)you get preview windows when you get ComfyUIManager and turn them on in settings then the sampler nodes have small liddle gen movies playing while you gen a picture
https://xcancel.com/drawthingsapp/status/1820881473429536774#m>Oh, one last thing. We have an experimental feature to allow you use negative prompt with FLUX.1 [dev] model.OH HE'S COOKING
>>101753828already possible
https://xcancel.com/drawthingsapp/status/1820881444241338436#mEven at 5bit it looks good, maybe the vramlets would be able to enjoy flux aswell
>>101753828>>101753854Maybe I won't kill myself
>>101753848>already possiblenot really, you can't remove blur out of a picture with the current negative prompts implementations for flux for example
>what do we want??!
>>101753854but unlike fp16 to fp8 there is noticeable loss in coherence with background elements. but still it probably outperforms sd 3 regardless. can you imagine? everyone prefers using 5bit instead of touching sd3 slop? kek
my new car
Are you guys also using the fp8 Clip Encoder? Or are you using the f16?
and finally my new business>>101753900>what do we want??!VRAM!>>101753985fp16 model, fp8 encoder .. unless the encoder shits itself and gets nothing right then I flip
>>101753985f16 with dev, as instructed by the comfyui tutorial page
>>101753985I use fp8
>>101753985fp8 T5 since there is literally no difference in quality and prompt adherence
>>101754017>>101753900Jensen is drinking our tears not gonna liehttps://www.youtube.com/watch?v=XDpDesU_0zo
>>101754110
>>101753983prooompt?
I think that we can use Flux to INDUCE the "divot tit" as a reference marker for another model, and I think we can apply weights around these markers, so so we can erase the work done by the other model.so in Flux, induce divot titsThen find the divotsThen (irrespective of the divots) use another model to add titsThen with a soft selection erase the addon layer, except in the tits.overlay the tit painting on Flux.
>>101754195there are adetailer models for breast detection already
>>101754177files.catbox.moe/dz22ho.png
>>101754214Great, so this isn't a real issue.
>>101754162
>>101754223Backrooms style?
>>101754231it is because there is more to NSFW than just nipples
>>101754255sorta
>>101754257I'm 100% canada safe.
>>101754257the whole body anatomy is rendered perfectly on flux though, so it's just a matter of adding penises and vagvag in there and we're good to go
I don't think we will ever get a 16ch sigma bros...
>>101754289>16chtalking about 16ch vae, is the flux's vae a 16ch?
I've been out of the loop for a long time(pre SD2). I've downloaded pony to satisfy my 1girls and am fairly happy with it, but for other stuff not so much. I know XL and 3 are nerfed in the people department, but are they generally good for abstract, landscapes, machines, etc? 25 it/s here so just "find out" takes a real long time
I just installed Automatic1111 and downloaded this popular Pony SDXL model plus the recommended VAE. I used one of the example image's prompts and parameters, but my results look complete ass.Picrel is the example I used.
>>101754220thanks fren
>>101754325Adetailer set to 2048x2048, lower cfg (~5.5), higher steps (~65), and get a style LoRA. Catbox it and I'll show you what I mean using the same prompts. Also, pony has a built in VAE. You don't need two.
>>101754325And this is my result, kek
>>101754388lol 512x512 .. probably default 20 steps? you need run higher res and probably atleast 30 steps
>>101754388you're not generating at the same resolution as the example image, anonie. XL is not a 512x512 model.And no, you don't need "at least 30 steps".
>>101754358>Also, pony has a built in VAE. You don't need two.It says "This checkpoint recommends a VAE, download and place it in the VAE folder." on its page, though.
>>101754437This is what I got without a vae and on autismmix DPO (the autismmixes of pony are better) with the same settings. I'll post one with my changes and show you what I mean.
>>101754436Ok, that's a good hint.>>101754489This looks great. Thank you.
>>101754223prompt? I keep getting windows in my indoor pools.
>>101754223is this flooox? i did a few backrooms gens but they were just okay
https://reddit.com/r/StableDiffusion/comments/1elqc3e/just_some_flux_images_it_pays_using_both_the_t5/looks like separating the prompts between clip and T5 really helps the model on understanding styles better
>>101754501>>101754489Here's the same thing except all I did was apply adetailer & cyancapsule LoRA at 0.75. I don't want to get b& for furry spam, so here's a catbox of the higher steps and lower cfg scale, still using the same prompt and resolution as the above.https://files.catbox.moe/0vditv.pngAgain, no VAE, just using autismmix DPO, a style LoRA, and adjusting the cfg and steps.
>>101754673Thank you
is FreeU worth including in your workflow (comfy) or is it just a schizo node?
>>101754699Freeu is placebo desu
>>101753854How do we remove the whorification weights baked into it?I am thinking negative prompts, but we'll see.It clearly finds a wide variety of faces and then tarts them up.
>>101754169Damn
>>101754685Np. Here's another one of their example images, but this time not only did I lower scale & raise steps and add a style LoRA, I also used some embeddings to help. Still using adetailer set to 2048 and autismmix DPO.https://civitai.com/models/332646/pony-pdxl-negative-embeddingsAnd catbox, in case you want to see the settings yourself. https://files.catbox.moe/20nrar.png
Don't prompt harder, prompt smarter.
>>101754554sorry lost prompt; >analog backrooms screengrab 1982 minimal white tiled pool with murky water spooky>>101754589yes fuux
Going full shojo
>>101754169prompt?
>>101754860Beautiful.
>>101753854So where do i get the 5bit? would actually probably prefer 6bit
>>101755041ty
>>101754220This is a very autistic workflow.
kek, I am redoing some old SD prompts on FLUX and I forgot to remove the BREAK keyword .. guess FLUX has a funny interpretation for that
>>101755186How do you prompt speech bubbles?
>>101755156thank you, I hoped somebody would notice
>>101755195as simple as it gets:>speech bubble: "Some text." works better with anime/cartoon, can work in 3D to.. better even is>some person says: "TEXT" for 3D, have it near the subject that is saying something or t5 doesnt get it
>>101755240thanks
Why loading 16 bit T5 and 8bit VAE separately yields worse outputs, especially with text than 8bit single file (Both unet and T5 at 8 bit?) from comfyui flux example page?I don't get it.Downloaded fp8 unet from https://huggingface.co/Kijai/flux-fp8 and everything else from comfy example pagehttps://files.catbox.moe/h6ahfr.pnghttps://files.catbox.moe/7a9jwl.png
>>101754231The entire I industry seems to be hellbent on having a single solution. Controlnet could fix a ton of issues that anon had, instead we have a new fanboy model of the week to fix issues with prompts not working. I can't count the number of times people don't want upscaling as a separate step and use substandard solutions because they would be forced to automate. Plenty of ways to inpaint boobs using a different model, nobody will bother sharing an implementation. Usually when I post fixes to these problems I get ignored.
>>101755209of course, it's quite beautiful.It's a shame that it fucks up nudes though, but as a final text pass it is a good flow. Pony is doing all the heavy lifting though.
Flux gives me hope for local gen. After Dall-e I felt pretty pessimistic for the future, didn't think we'd get anything close to that quality. But goddamn Flux is close
>>101755209Nice workflow, surprisingly fast on a 4090.Would the opposite not work better though? Flux for composition with a 2nd pass using a Pony-based model?
>>101754437I test every time and document it. I have had too many issues trusting docs and idiots saying vaes are always included instead of saying the are mostly included.
>>101755570This rules.
prompt: marquise cutok, unusable, deleting immediately
>>101755646thank you
>>101755496indeed it does, I was just curious if I could get sensible text working in an inpainting pass with flux. I'm new to comfy, do you maybe know any good nipple detection?>>101755556pony fucks up any text. maybe text detection for masking could be possible.
>>101755052>>101753854any idea where to download that 5bit?
why does it even add the weird nipple censor, there was nothing in the source image
>>101755021>https://litter.catbox.moe/jui298.pngwildcards made with chatgpt
>>101755783Probably because there are some images present in it's training to know to put something there, but not enough to make it do it correctly.
>>101755795This is so great I can't even
>>101755855omg fr i love this so much. appreciate you anon! ^_~
>>101755872NP :)
>>101753827>ComfyUIManagerty, not sure why it doesn't see my already added (sideloaded?) model, but whatever.
>>101755941lol
>>101755798>1.5Gay
>>101755897Prompt?
>>101755696I have a nipple workflow somewhere, used it for SD3. Could easily be adapted for Flux. Only works on breasts though, so you still only do erotica, but it's something. Let me find it
>>101756022started with this prompt (SD 2.1 image attached to compare)>18mm photo product aesthetic microcomputer made of pastel plastic transparent 2000s 1977 (home one MC80 Star Cruiser:1.2) (Tantive IV) detailed electronics (empire strikes back:1.4)(transparent clear:1.41) hybrid mini biological computer micro with (small crt screen) marble 1980 table chunky (child's toy) [glass] (jelly:1.22) (ombre:1.2) (toy) (transparent:1.4) (clear plastic) flcl, cinematic 18mm (photo) vintage movie still (flesh meat wearing) biomechanical
>>101756056Thank you
>>101756009ur mom is lol
>>101756183Like from old stop-motion movie, looks great
In ComfyUI Manager, I don't see the checkpoint I downloaded, but I used Refresh.
>>101755696>>101756024https://files.catbox.moe/t52e16.pngThis is it. Pre-adapted to flux. Used 1.5 for nips though but could easily be adapted to Pony.
Being able to actually get an image based on your prompt from a local model is a gamechanger.
>>101756301sweet, thank you anon
>>101755982cool it with the interracial stuff
>>101756378getting too horny?
>>101756308STOP THAT INTERRACIALISM!no more hot chick with black men!
>>101755502Flux is already better for many purposes, such as text.
>>101756478It's fine, she's a Latina and although Shadow is black, he is also an emo, so they both kinda meet in the middle, racially.
>>1017560161.5 (exquisite details) will always feel like home, even if it's janky.
So does anyone know why the Flux speed fluctuates so much?
>>101756692Like 1 second I am getting really fast gens, like 1.3 sec/it on dev (no fp8) on my 3090, another that it goes down to 7 sec/it, and then I have to switch to fp8 which also does 1.3 sec, which also sometimes goes back down to 3 secs or 5 secs, and it makes no sense the fastest I've seen is 1.3 sec it does not go any faster.
>>101756625Yeah it's great and I still use it. Have you ever tried training loras with it? Who even made it and with what dataset
>>101756721Found the issue, when it does that it says "loading in lowvram mode 21981.199919891358"
>>101756301adapted it for flux and ponyfiles.catbox.moe/owz6da.png
>>101756794cat box plese
>>101756794not beating the allegations ani
repost is real
>>101756794No give her a massage
>>101756796https://files.catbox.moe/9ck5wm.png
any tips for flux inpainting?
>>101756837>NoNow*
>>101756478What about this kind of interracial?
>>101756794>>101756840based
>>101756778Better than my quick and dirty conversion. Thanks
it big
>>101756905The fabled CRT game boy
>>101756916the turbo express
>>101756905cool
>>101756794Leave the fox alone comfyanon
>>101756865it detects the dressed boobs too, too bad! I wish the skin detectors would work for cartoonish styles
>avatarfag gets most (You)s again
>>101756794>>101756840you're shit (as a human)
>>101757140ToT
>>101757140>>>/b/degen
>>101757140
have a nice car
>>101757154>>101757160>>101757187this model is too safe for that
>>101757195Yet there is cunny, plain as day.
>>101757198are you saying the model is based?
>>101757140incredible stuff
ever since i updated comfy my gens have been a lot slowercomfypls
>>101756963likewise
>>101757248sheeit dat ncie
>>101757195the point was that you belong to >>>/b/degen and not>this model is too safe for that
>>101755078love this
>>101757279new to /g/?
I came to flux loli
So has anyone gotten the model to do a middle finger? Is it just a skill issue?
please moot save us from this cunny
>>101757315your parents must be very proud
>>101753346
it just passed midnight
>>101757024I have figured it out.files.catbox.moe/ery3a6.png
>>101757581Thanks Anon, I will check it out after work. Safe gens only for a few hours.
isn't there a seed node that I can feed to other nodes?
I just want to say I love diffusion. It makes my Photoshop projects fun and easy, I don't have to google some stupid stock images. The quality of life has improved greatly.
>>101757317I thought I'd be witty and nail it but it is harder than expected>>101757336mootykins is gone
Does anyone want to make a discord for /ldg/?
how THE FUCK do I make a seed node?
>>101757492At least gen it, Anon.
>>101757745it's a primitive. double click the seed input
>>101757317yes but by accident
>>101757745>>101757768no wait it's not. what did the manager say? there really isn't much of a difference to a primitive other than it can reroute
>>101755897flux keyboards are so gooooooood
>>101757745How do you make the connectors more autistic and straight like that? It looks nice.
>>101757768>>101757795at least that worked, it was driving me crazy>>101757816dunno lol, saved the image from github
>>101757925She looks like she is very small.
>>101757983>The scene depicts a woman heading to the bank to deposit her onlyfans earnings in an offshore account to avoid taxation.
>>101757317i got one after many failed attempts, it generally does index fingers instead
>>101758036forgot image
>>101757907nice
>>101757907>>101758059Really cool
>>101758056those are impractical shoes for a heist
>>101758059>This image depicts a ghost angrily leaving her room and throwing her ghost dildo at the viewer
>>101758098That's a fucking moby huge
>>101758059>>101758105How do you think she died, Anon?
>>101758117lmao
How do I get BIG bobs in flux?
that's a good effort
>>101758130>Making fake slides for presentation. Cute idea
>>101758140What happens if she farts? Does it just create a bubble that slides around the suit?
>>101758166Her asshole inhales the fart as soon as possible
>>101758148
any gradio demo? link down
did it milla jovavich her?
So what would make Flux perfect right now?
>>101758241Porn.
>>101758241finetunes, loraseverything would spawn from that
>>101758241make it run quicker you fucking retard
>>101758241Same model but smaller. This model is so big that there won't be many good finetunes. The cost of actually finding the right settings will be insane
>>101758241Not requiring hardware from 2034.
>>101758353hentai x ray gone too far
>>101758241the styles, aesthetic creativity, and sovl of october dall-e
Immaculate flux gens ITT
>>101758312
Come and get your next loaf of bread...>>101758249>>101758249>>101758249
>>101758241>inpainting support>controlnets>style/face/whatever transfer adaptersThat's the bare minimum to get it usable in saner pipelines like Krita Diffusion and make what you want, not what this thing wants.Tooling should always come first. Too bad they are AI eggheads and don't understand that. Either a "funny picture generator" is their actual target, or they are genuinely misguided enough to believe that text prompt is enough.
>>101758283Is that supposed to be like Makoto?
Flux is interesting.Text is hard to define when the model doesn't know what you're talking about as to where to put the text I guess.
>>101758312It doesn't?I don't get people complaining about, it works fine on a 24g card
>>101758803Very few people are running on FP16, but it is definitely better at text. I just can't wait 20 minutes per image.
>>101755106>>101754860Would you catbox either of these anon? They look great
>>101758241actually being on par with based soulful de3 prompt understanding
I've been lurking this thread since flux came out and nobody has made a single pepe with it. bros...
>>101760556i tried a few prompts with pepe in them and it ignores "pepe" or it makes a generic cartoon frog. it's joever
>>101758760I honestly don't understand doing text purely on ai. Literally the worst tool in the toolbox for it, but I guess since there's poor gimp integration (any?) people feel disconnected.
>>101761790We need to fund froggen.
