Discussion of free and open source text-to-image modelsPrevious /ldg/ bread : >>101917856>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/ComfyUIForge: https://github.com/lllyasviel/stable-diffusion-webui-forgeInvokeAI: 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>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>Maintain thread qualityhttps://rentry.org/debo>Related boards>>>/h/hdg>>>/e/edg>>>/d/ddg>>>/b/degen>>>/vt/vtai>>>/aco/sdg>>>/trash/sdg
i think ai images SUCK
>>101920566AI thinks you suck, Anon.
>>101920549NTA but last night I went to bed at 2 AM trying to figure this out. I think something's borked with Forge's "automatic memory management".
God I hope Comfy implements better support for multiple GPUs Its annoying as fuck to fiddle with the workflow for loading LORAs
I spent the past 5 hours generating and fixing boomer prompts for my lora dataset and I'm not even a quarter through all the images I have 500+ datasets with anywhere from 80 to 1000 images in each. Save me ldg my autistic hyper focus won't let me stop, I'm scared for my future
>>101920607Only if you consider memory leaks malware.
internvl2 sucks dick i fucking hate benchmarks
>>101920615>AI thinksthink again
>>101920686>chinks lying and cheating their way to "success"NO WAY!
>>101920566>>101920699Apparently Impact is a "Striking, dramatic font"
OH IT DEFAULTS TO A TINY MODEL, WRETCHED SETTINGS, AND A CHINESE FUCKING PROMPTMAYBE THAT HAS SOMETHING TO DO WITH IT BUT I CAN'T CHECK BECAUSE IT ERRORS WHENEVER I TRY TO USE THE 26B
>>101920766are you ok anon? :(
>>101920781i want good thing and world keep giving me bad thing instead
so Q8 model is like fp16 but more efficient?
>>101920444autocaption as comparison... https://github.com/Z-L-D/Autocaption
>>101920686Alright but you are judging 26B which is the worst open version they have, not 76B or Pro version.
>>101920820dude that's a 7B he's comparing it to
>>101920808man that's limp>>101920820no i was judging 8B because that's what the space defaulted to and it errors whenever I want to try 26Bbut joycaption>>101920414>>101920444is also 8B and it blows that shit out of the water
>>101920808Complete, concise, exhaustive. Local won.
>>101920823See >>101920808In that case it was 8B, even worse than 26B
>>101920636i see. yea, i'm seeing a lot of other people are having the same issue, which is somewhat comforting. hopefully illyasviel finds a fix for it. for now, back to pony/SDXL i guess.
that lad who got 1.5k steps of lora training on 8gb vram in under 5 hours, how did it go?
>>101920851died in a housefire
Flux can't do sweaty women like DALL-E 3 can
>>101920849I think Forge is a prime example of a hobbyist developer who goes parabolic, churns out a ton of low quality code really fast, and the project ends up grinding to a halt or collapsing after the technical debt gets too large. Many such cases.>>101920869Or like SD 1.5 can.
have they finetuned soul into flux yet?
>>101920869I couldn't get it to do any real degree of moisture
>>1019208761.5 can't really
>>101920851I think you might mean me (I don't know, if there's some other vramlet anon who posted about this I'm sorry). I didn't test the loras since they had no boomer prompt captioning, working on that then will run training again.
>>101920895what were your settings?
Blessed thread of frenship
>>101920876indeed
What comfyui workflow do you guys use for flux?
THREAD THEME: make something cool in the sky
>>101920900https://desuarchive.org/g/thread/101908455/#q101910682based on the low vram use I saw I'm guessing I could try to push the settings a bit more, that was just based off kohya's recommendations for vramlets
>>101920935thanks, i'm gonna give it a stab tomorrow too
>>101920869https://civitai.com/models/642235/sweaty-shirt-armpit-sweat-pit-stains-wet-spots-flux-dev1
>>101920880For anime these LoRAs look interesting, everything else mehhttps://civitai.com/models/649031/inksketch-flux?modelVersionId=726131https://civitai.com/models/640405/flux1-dev-modern-anime?modelVersionId=719943https://civitai.com/models/647940/flux-atilessence-lora-test?modelVersionId=724910https://civitai.com/models/648623/ascii-art-flux?modelVersionId=725667
>>101920869all bodily secretions or excretions are forbidden.
>>101920907Dual CLIP loader for text encoders.Force model to device to force CLIP and VAE to CPU.Bootleg GGUF loader to load Q4_0.VAE loader for ae.safetensors.Then everything else as normal.But it's really slow. I'm sticking to 1.5 until I get a new GPU.>>101920931I'm at work REEEE
>>101920946>skin remains perfectly dry
>>101920941good luck anon, may the vramlet gods smile down upon us both
>>101920957Thanks
>>101920959>>101920950>>101920904>>101920869Works on my machine
>>101921001Her shoulder looks like a salted pretzel
>>101921001mpox-coded
>>101921001it's the fucking monkeypox
>>101921001bitch lookin like a hamburger bun egg wash and sesame seed ass ho nigga
>>101921001she has fungus growing on her
>>101921001Why does flux love the shitjourney wax look?
you guys just ended his whole genning career
>>101920869dalle is king>>101921001looks...bad
>>101921001congratulations, you made flux look like shit
>>101921091but anyway.. not a dalle thread
Can't believe I'm saying this but I'm bored of booba. Time to explore the latent space for gen ideas.
>>101921008>>101921009>>101921010>>101921016>>101921031>>101921032>>101921091>>101921093kek, what about that one?
>>101921164god man, call me back once flux can gen actual humans and not one of those stupid fucking realistic wax figures. fucking SHIT! i HATE THIS! FUCK!
When is the LoRA to de-chin butt coming out?
Anyone here making LoRAs yet? I can't get over how they come out almost perfect all the time.
Despite the occasional meme chairman Trump gen I'm starting to see how Flux is visually stunning, but really limited.
>>101921192I would love to. I made a bunch of LoRAs for XL, would love to re-do them for flux. I am smooth brained with a beefy gpu though so need to be spoonfed a training script for Kohya.
>>101921174refreshing>>101921182she seems to like her job
>>101921219https://github.com/ostris/ai-toolkitIt cannot get more simple than this. In fact, I found it simpler than Kohya. I will switch when Kohya is fully cooking again though.
>>101921174god I wished I could get the Loras working
>>101920270The thing is I'm still able to run original dev/schnell model with fp8 dweight. Even with t5 fp16 it still running without OOM despite I only have 16 gb ram. It's probably just nf4/gguf nodes implementation problem.
>>101921100>>101921091kek
>>101921192It's crazy how well the model generalizes. You can include a dozen well-captioned nudity shots in the dataset and it would just work. In early SDXL days restoring anatomy knowledge was way way harder.
>>101920567I get "AssertionError: You do not have CLIP state dict!" error and if I pick clip_l in the VAE options I get "AssertionError: You do not have T5 state dict!" What else should I add to make the flux Q5_1 work in forge?
>>101921241I am just used to kohya, my problem is I am a retard with the settings, I don't know what to set any of the shit to, so that's the bit I need spoonfed.
>>101921289The future is looking bright. I didn't even need to caption my shit and it restored anatomy.
>>101921164lower your cfg to 1.5 - 2.0 and try again3.5 gives too much of a waxy feeladd some additional to the promptalso i'm pretty sure that negative prompt doesn't do anything because flux
>>101921306>also i'm pretty sure that negative prompt doesn't do anything because fluxit does, it removed the sand she had on her shoulder on the first render >>101921001it can even remove blur if you want https://imgsli.com/Mjg3OTYw
Are loras working with gguf on comfy?
>>101921330That is a man
>>101921338Here, now you don't need to ask anymore. https://civitai.com/search/models?sortBy=models_v9&query=gguf
What is the difference between this place and /sdg/? They're both primarily Flux threads.
>>101921365Didn't even answer the guy's question and linked to some checkpoints the guy probably already has. Useless retard.>>101921338Not as of yet.
>>101921383Go to /sdg/ and ctrlf "PW" and see why we avoid it like the cyst it is.
>>101921362nobody asked
>>101921306I went for GuidancePos = 3, going lower gave me weird artifacts, so far that's the best one I got
>>101920268how "vintage" are we talking?
>>101921402>Didn't even answer the guy's question and linked to some checkpoints the guy probably already has. Useless retard.I linked him to the place where LoRAs will appear for gguf when they start appearing numbnuts, so he doesn't have to ask here literally every thread like he has been doing.
>>101921451>where LoRAs will appear for ggufIt's very likely that LoRAs can be used without the being specifically for GGUF, stop jumping the gun. You have fucking downs syndrome and nobody has the heart to tell you you fucking retard.
>>101921430your multiple sweat prompts caused it. it's doing what you told it to.
>>101921468First of all, it's Down syndrome, not Downs syndrome, and you should know that. Also speaking of Down syndrome, it's impossible for flux to generate a person with it. We need a retard LoRA.
>>101921508really? it can't generate mongolians?
>>101921231Of course, it's pretty cushy just triaging to the people who do the real work and have to actually take their job seriously
>>101921570Maybe there's some magical prompt that can do it but I haven't found it and it sucks because I want to make tard memes.
Can't get comfyui to run with zluda on windows
the aftermath of yjk
>>101921496Ok, fair enough. I tried it again and used the word "sweat" only once here
>>101921647I look like that
>>101921658Rich pipo looklike that
>>101921668In my country poor people look like that. Rich people can afford vegetables.
>>101921658Can you see your own dick ?
>>101921680In my country poor people eat vegetable and rich people can afford meat
>>101921455Prompt for this aesthatic ?
>>101921696The physique is achieved by eating exclusively fried chicken and burgers because they cost as much as a single capsicum.
>>101921716That's just the default look, I threw in some song lyrics.
>>101921612How DARE you
This says a lot about society
>>101921642looks 100% better. drop the cfg even more, try 2.0
>>101921817This thread is for AI images only, no real photos allowed.
>>101921812by "cfg" you mean the actual "cfg" or the GuidancePositive?
>>101921828can't you see the 'AI earrings' ?
>>101921874I don't know Anon, my wife has all kinds of retarded earrings, I can believe them.
>>101921901LUV ME NODESLUV ME SPAGHETTILUV ME 1GIRLSATE WEBUIATE GRADIOSIMPLE AS
>>101921908I look at too many AI images kek
This is harder than it should be. She is literally in water and bone dry
>>101921944what are you trying to achieve?
>>101921957wetnessI did manage to achieve a consistent person though, but I want to make a person wet. Flux can't into wet people.
>>101921977>Flux can't into wet people.yes it can, just look at a few posts higher >>101921824>>101921642>>101921430
>>101921855not sure what it is in comfy, but it looks like what you changed last time (to 3.0), so probably that one.
>>101921988I am not trying to achieve a sweaty person, I am trying to achieve a wet person. This is the closest so far. Gonna try dropping the guidance, see if it makes a difference.
>>101922008give me your prompt and I'll see what I can achieve
>>101921824>>101922002>hair is dry
>>101922023noice
>>101921989>not sure what it is in comfy, but it looks like what you changed last time (to 3.0), so probably that one.it's the GuidancePositive, it's the equivalent to "distilled cfg" on Forge, and like I said I can't put this lower I got fucked up results, this is my limit
>>101922024
>>101921988>>101922022Nta but you cant get it i think tried tried every word for wet still didnt get it and if you at it try getting some soap bubble too
>>101922022A woman standing inside a waterfall with water pouring on her head and over her body, her skin is shiny and glistening from the water, and her hair is dripping with water.
>>101922080>Nta but you cant get it i think tried tried every word for wet still didnt get it and if you at it try getting some soap bubble tooIf you haven't tried Tonemap then it means you haven't tried enoughhttps://reddit.com/r/StableDiffusion/comments/1estj69/remove_the_blur_on_photos_with_tonemap_an/
>>101922070She has a hydrophobic coating
what is the status of finetune rn?my dataset is waiting for the 8xh100 :>
>>101921828Sorry, won't happen again.
anyone going deis/ddim_uniform? getting good results with it>>101922099does>her skin shines, slick with a thin layer of water.help your gen?
>>101922086This is my first attempt, I think I can do better
>>101922086>>101922164ok that one is good, never touching GuidancePos ever again kek
>>101922153intersting. I used "a thin layer of crystal clear gel"
>>101922200goddamn that looks good
>>101922203It would be better with real nipples. https://files.catbox.moe/hvrecd.png
>>101922203wut lol no it doesn't, the got the mpox or some pond scum on her, her skin looks like wax, and the edges of the tan lines look way too sharp.
>>101922220just give it a low-denoise img2img pass in one of the SDXL models.
>>101922093>you didnt tried this trickYou also impaint it using sdxl models
>>101921828anon... they literally have the ai same face
>>101922153>her skin shines, slick with a thin layer of water.Helps a little>>101922238I have a full SDXL nippling workflow for flux. But in this case I am not interested in the nudity, not even interested in the woman really, I just don't like when I have a hard time achieving something, especially when the model is so capable in most other areas. It's a personal challenge to make a wet person now. https://files.catbox.moe/h8avmw.png
>>101922255why should I need inpaint, Flux can do it on its own
>>101921231brap
>>101922305Just looks crispy
>Change strength of LoRA in workflow>Entire model reloads on to vram from scratchWhat the fuck is causing that?
>>101922261give me the workflow of your picture, I'll try to improve it
Pro can do it. Dev limitation.
>>101922473posted >>101922220
>>101922447Idiotic devs
>>101922220bruh your workflow doesn't have a seed how can I replicate that in the first place?
>>101922447Same issue on sd.next
>>101922260They’re siblings, you bigot.
>>101922476Pro vs Dev
>>101922533Now that dev has functional and extremely effective LoRA support and at home fine tunes on the way, I really won't give any shits for a few months until this model has truly been tapped
>>101922533other way around, or is Dev actually better?
>>101922545wait im retarded and can't into read. esl sirs must understand
>>101922545nta, but I think Dev is objectively better like 60% of the time. Oftentimes it's a complete tossup though.
>>101922533desu dev can be closer to pro with CFGmaxxing and finetunes will improve this shit further, the future is looking great
>>101922540I agree, I just hate hitting these kinds of limitations. I know for sure someone will do a wet LoRA. But it's less about the actual wetness itself and more the fact that I hit a limitation. Still, nothing like SD3 where day 1 we were struggling with a woman laying on grass. >>101922545Pro is left Dev is right how can you think right looks better? Left is clearly wetter.
>>101922570sorry sirs>>101922551
average spanish tourist trap
>>101922447Comfy should hang for this bullshit. We have had enough, bros.
>>101922598Beautiful. Would live there.
>>101922586
Anyone tried this?https://www.reddit.com/r/StableDiffusion/comments/1et0mlj/guide_to_use_flux_on_forge_with_amd_gpus_v20/
>>101922622why is she holding the sign backwards? is she stupid?
>compose with flux>upscale with SDXL>it fucks up the handsevery fucking time
>>101922664denoise is too high matey
>>101922656>tourists go home>refuge welcome>is she stupidyes
It's true that flux is "censored" ?
>>101922683nah mate look at these nips, wheyyy
>>101922664Don't upscale with SDXL. Just alter specific parts with masks.
>>101922683I doubt there was any porn in the training data, so I would say no
>>101922691>can't make a wet woman>can make a wet obese bald manwhat did they mean by this?
>>101922691>>101922694Not porn but controversial prompts. Can it do what grok does?
>>101922683Depends on what you mean by censored.
>>101922693The problem is Flux looks kind of fuzzy and low resolution with that awful DoF so I have to upscale one way or another. Perhaps its because I'm using low VRAM version
>>101922713Mate. Grok IS Flux.https://x.ai/blog/grok-2>In collaboration with Black Forest Labs, we are experimenting with their FLUX.1 model to expand Grok’s capabilities on X
>>101922683Depends what you mean by censored. It appears to have like 0 pornographic material in the training data, but there were clearly some tasteful nudes in the set. Even the most basic of LoRA training puts high quality nudity right back in too. So it's not really censored so much as lacking excessive NSFW content. But model seems more than happy to learn NSFW stuff in short order.
>>101922713You're gonna get a warning for that. I did one of Trump shooting up a school and got the janny stink on me, all the while some retard was posting literal cp.
>>101922713>Can it do what grok does?
Give me the low down on Grok? I don't think anything could make me subscribe to a phone app. I don't login to anything ever, rather suffer with janky ComfyUI and waste hours of my time.Elon's gonna make it open source soon and then we will all be making hardcore Grok gens with ComfyUI... right? Gens we've never dreamed of. I CAN'T WAIT
>>101922731I bet SAI are fucking FUMING
>>101922736I deleted the file but like that>>101922731Interesting
>>101922772Bait used to be believable.
>>101922772It's literally just a straight dump of Flux.1 Pro with no guard rails.
>>101922772
>>101922773They will be fine. They are trying to make SD extra safe. When they achieve it, they will be raking infinite money. The AI game is all about safety.
>>101922784kinda goes hard lad ngl
>>101922772To obvious, but it's pretty astounding the amount of people I've legitimately compare flux with grok and not realize they're basically comparing the same model.
>>101922810isn't it weird how bing is just as good as dalle3
>>101922814Totally mind blowing.
>>101922664Thanks for the prompt idea Anon
I think what's interesting about BFL is that they've actually managed to make a model that front ends can service to their customers and people actually like to use. SAI was never able to do that. Their API was only ever used begrudgingly in the absence of the open weights.
>cfg 1>"A sweaty woman">Gives me a woman with blood on her facewtf flux?
>>101922839how can flux think sweat is blood?
>>101922810I really like the fact that Grok and Musk are getting blamed for all the edgy gens and not Flux kek
>>101922848Maybe she has hippo dnahttps://en.wikipedia.org/wiki/Hipposudoric_acid
>>101922848is there a point to doing 30 steps instead of 20?
>>101922868They're mainly seething because twitter released a competent product.
>>101922884>is there a point to doing 30 steps instead of 20?it's for more consistency in qualityhttps://reddit.com/r/StableDiffusion/comments/1er3wt7/if_you_want_a_good_compromise_between_quality_and/
>>101922848We need better datasets, man. Can't help but feel we aren't even close of taking advantage of the 12B parameters.
>>101922772It still has limited instances, can't generate unlimited images>>101922848How much time it takes to generate it?
>>101922902>How much time it takes to generate it?
>>101922810They are clearly different models. Maybe based on the same underlying architecture, but they give similar yet different outputs Here is a comparative I made.https://www.reddit.com/r/StableDiffusion/comments/1etjfbr/flux_reply_to_grok_2/
>>101922800>The AI game is all about safety
>>101922943they probably finetuned pro to make it better?
>>101922943>They are clearly different models.I am not sure they are. Aren't user prompts fed through grok then to BFL? It's possible the prompts grok feeds to the API are not the same ones the user typed in.I'd imagine they just use flux pro. Is it stated anywhere they have any other models?
https://civitai.com/models/650637/liarinchief?modelVersionId=727931
>>101923022a Lora trump? lol, flux can render Trump just fine on its own
>>101923022>Dude is just sitting there shouting at phantoms about how he won't back down to bullying.
>>101923036Why are you defending the liar in chief, bootlicker? Is that boot tasting good?
forge dev is a fucking numbskulltoo lazy to scroll up, are other people complaining about bugs in the latest forge too?
>>101923022>Discussion is turned off for this model.just wanted to praise his choice of a sensible rank size
>>101922943>>101922893>>101922641>>101922093omg, you do reddit? THE NARWHALS BACON AT MIDNIGHT! haha, upvotes to the left fellas.YOU SIRS, WIN THE INTERNET ON THIS BLESSED DAY
>>101923070Imagine siting there waiting 7 minutes just to see this.
>>101923095shit i forgot a line of my post this is forge failing to generate a grid
>>101923070>filenamekek
>>101923067>This is the guy calling you Liar in Chief.
>>101923121lmaooo, they really all have the same face isn't it?
>>101923022
>>101922893all that shows is a cartoon miku examplei think for realistic looking people 20 steps looks more natural, but that might depend on the rest of your setup too ofc
>>101923177
>>101923232
if you guys are gonna post trump, do "trump doing Gangnam Style on the lawn of the whitehouse" >(improve prompt as you will)
>update forge >now VAE's don't show upwhat is this guy doing? really?
>I can't wait for pony dev to make pony flux>150 upvotes
>>101923300What's the workflow for this? All I get is slop on flux comfyui, takes 7 minutes to generate a 1024x1024 too
>>101923345>flux>cfg 8>karrascome on now
>>101920957>slowQ4_0 and Q8_0 are really slow somehow but the dev_fp8 version works flawlessly with some adjustment of course, generation time can vary between 1:30min~5:30min
>>101923367Fantastic gen, Anon
>>101922045Is that dalle or this:https://civitai.com/models/646686/japanese-photo-1980s-style-1980?
>>101923394How do you get it down from 5 to 1 minute? Nothing I do on my 3060 seems to work.
>>101923394>generation time can vary between 1:30min~5:30minIt offloads to your RAM or, even worse, pagefile.
>>101920931I missed this. I miss thread themes too. Next time post this with an example gen of the theme you want to see.
>>101920553WHAT THE FUCK THERE ISNT ANY FOCER/SET CLIP DEVICE NODE I CHECKED EVERYTTHING FUCKING DISGRACE PIECE OF SHIT
https://civitai.com/models/651004/giger?modelVersionId=728323Finally some actual kino
>>101923490>what is a custom node
>>101923345I'm on forge, it just works
>>101921189prompt japanese or korean, they don't suffer from it
>>101923490it's here https://reddit.com/r/StableDiffusion/comments/1el79h3/flux_can_be_run_on_a_multigpu_configuration/
>>101923490ComfyBootlegOffload.py
>>101923536I don't like Asian people
>>101923490>https://www.reddit.com/r/StableDiffusion/comments/1el79h3/flux_can_be_run_on_a_multigpu_configuration/
>>101923345>flux>karras>cfg > 1>negative promptWhat are you doing?>>101923526Forge is slower and buggier than any other UI out there.
>>101923345>>101923561Worst of all it's in light mode
>>101923345you can't put a CFG > 1 without adding or else Dynamic Thresholding or Tonemap, what are you doing anon?https://reddit.com/r/StableDiffusion/comments/1ekgiw6/heres_a_hack_to_make_flux_better_at_prompt/https://reddit.com/r/StableDiffusion/comments/1estj69/remove_the_blur_on_photos_with_tonemap_an/
link to the qf4_0 flux model?
>>101923345Since nobody actually explained and just gave you shit, Flux only works on cfg of 1, and because of this negative prompting does not work. Set your cfg to 1.Remove your negative prompt text (keep the node)Change your sampler to "Simple"Run again, it will be faster and good.
>>101923467>Next time post this with an example gen of the theme you want to see.i'm making it vague on purpose so people can interpret it however they like
>>101923606https://huggingface.co/city96/FLUX.1-dev-gguf/tree/main
>>101923540THAT ISN'T A FUCKING WEBSITE YOU PIECE OF SHIT JUST SEND ME A FUCKING DOWNLOAD URL OR WHATEVERFOLLOWED A YOUTUBE TUTORIAL FOR 4 HOURS JUST FOR THIS SHIT TO NOT WORK OUT OF THE BOXWHO WOULD DESIGN IT LIKE THIS. I JUST WANT TO DOWNLOAD IT WITHOUT HAVING TO DO THIS WEIRD SHIT, WHY ISNT THERE JUST A .EXE I CAN INSTALL??
>>101923616I usually only read prompts that contain gens.
>>101923618
>>101923614he can put a cfg > 1 and use negative prompt if he uses DT or Tonemap though
>>101923614sgm uniform is the best scheduler
>>101923629about to try comfyui for the first time in ages, please catbox this so i can replicate it (nude) (this made me horny)
Any recommended optimizations for the 3060 please?I am stuck with 5 minutes per gen no mater what I do, redownloaded the models, reinstalled comfyui, updated the gpu drivers but nothing works
>>101922598Home...
What UI do I need to use to use flux GGUF? It doesn't work on comfyui
>dark theme
>>101923657Buy a better GPU.>>101923665Are you a zoomer? Literally use a search engine.
>>101923665https://github.com/city96/ComfyUI-GGUFcome home white man
>>101923657Force the text encoder to your cpu with thishttps://www.reddit.com/r/StableDiffusion/comments/1el79h3/flux_can_be_run_on_a_multigpu_configuration/And with your 12gb of vram, run a model that asks for less than 12gb, go for Q4_0 or Q4_1 for example
>>101923685go back to your containment thread d*bo
>try a workflow after several minutes of searching for a basic fucking txt2img because comfyui users are linux adhd spergs that need 20 billion controls>run it>realize this bitch is doing 2 extra levels of upscaling in such a way that it blew up a 1536 image to twice the size>didnt even run out of vramokay, credit were its due, that's kinda rad. But how do i set this up so i can CHOOSE if i want gens to be upscaled?
>>101923698It's amusing how every general on this site has a resident boogeyman, and everyone I don't like is that boogeyman.
>>101923635He is already retarded please don't confuse him. Keep it simple
>>101923706shit my bad i thought i linked this at the endhttps://comfyworkflows.com/workflows/e6c1c436-f878-4cc3-be0a-43ee96864467>>101923711to be fair it is true that every general has at minimum 1 schizo
>>101923713true true, fair enough
>>101923697I did that already, forced the clip to cpu and for the model I am using flewx1-dev-fp8 it's what I saw users reporting 2 minutes or so per gen by using it.
>>101923721That's a bit overkill. Why not just do this? >>101920957
>>101923757brother i have no clue what's overkill, this shit isn't intuitive and there's no real handholding, download a preset workflow and pray it works.also im not even using flux its just sdxl.i want to try flux after i get the hang of comfy again though.
>>101923743>I am using flewx1-dev-fp8 it's what I saw users reporting 2 minutes or so per gen by using it.fp8 asks for 14/15 gb of vram during inference, I'm actually running it right now, piercel pictureYour 3060 isn't enough, it's offloading the 2/3 surplus of memory to your ram, and that shit makes everything slower
>>101923628 oookayyyTHREAD THEME: delicious cats
>>101923777Wait, seriously? The Kijai one?
Is there any way to fit q4_0 in 8 VRAM? I'm already putting clip and vae on the CPU, but Comfy still gives me OOM errors unless I go below 512x512, and it's slow as hell anyway.
>>101923789I hope we get a finetune that fixes the Madame Tussauds syndrome.
>>101923818yes
>>101923821did you activate the system fallback on Nvdia Control Panel?
>>101923789She's got MY vote.
>>101923841I'm using AMD on Linux.>inb4 amd
>>101923391>>101923561Saw others use cfg 8>>101923526I'll try forge next, takes too much to generate now>>101923587>>101923614Can flux do anything other than 512x512 and 1024x1024 properly?
>>101923850and MY axe!
>>101923853>Can flux do anything other than 512x512 and 1024x1024 properly?yes, it's really resiliant to resolution changes, you can have fun with it without much issues
>>101923821Werks for me. Restart your pc, turn off all the bloatware and try aga->>101923851Nevermind, my condolences anon.
>>101923853Yes
>>101923825I will try the Q4_0 then, I will seriously go crazy if this turns to be true.
>>101923853This is better but still has that slop look compared to 1.5
>>101923851>AMD>on LINUXquit while you're ahead
Can flux be uncucked?When will the nsfw models come out?
>>101923878go for Q4_1, it also fit to your VRAM and is a better quality
>>101923878Try the 5 or 4_1 quant, they should fit just fine
Bread so fresh you can smell it...>>101923884>>101923884>>101923884
>>101923888I am NOT ahead.
>>101923853>Saw others use cfg 8they use Dynamic Thresholding or Tonemap to make it work >>101923587
>>101923853their paper says up to 2mp, i've done 1408x1408 fine as well as higher res common aspect ratios like 4:3 1152x1536. Just keep it divisible by 8
>>101923853>>101923877And yes again>Can flux do anything other than 512x512 and 1024x1024 properly?
>>101923894>>>/h/hdg>>>/e/edg>>>/d/ddg>>>/b/degen>>>/vt/vtai>>>/aco/sdg>>>/trash/sdg
>>101923886You can't un-slop schnell, try dev
>>101923928Those niggers only want to post their slop. Not talk about technical stuff. Nigger.
>>101923945>bu..but...
>>101923539>>101923540>>101923559Thank you, nice white men>>101923514Fuck you
>>101923918Takes too long to generate 1024>>101923929That's slower
>>101923977>t5xxl_fp8
>>101923990stop using those fucking brackets in the promptit won't solve anything but they annoy me
>>101923990>>Takes too long to generate 1024I have a 4080, but I noticed my it/s didn't really change as much as you'd expect from jumping the resolution. Maybe like 20 seconds longer per gen, these take me about 75 seconds a gen for 30 steps
>>101923993>sir the t5xxl_fp16 tits you orderedflux1-dev-F-16.gguf crashes. should i load the t5 on the cpu? is that it? i hate being a 32ramlet so much
>>101924128>should i load the t5 on the cpu?yes you should do that
>>101924128>should i load the t5 on the cpu? is that it?yes
>>101924128>>101924143>>101924148is that not what's already happening in the screenshot with the Force/Set CLIP Device node?
>>101924159and which device is set in that node, retard anonie?
>>101924159>cuda: 0Anon, your cpu/ram are not a cuda device.
>>101924159no, it's putting your text encoder to "cuda:0" a.k.a your gpu, you must change this value to "cpu"
>>101924143>>101924148>>101924159>>101924165>>101924167Still crashed :,,-(
>>101924199the fp16 is just too big, go for Q8_0, it has the same quality while being 2 times lighter (12gb) >>101923697
>>101924199Is it litearlly just impossible? to t5xxl_fp16 and flux1-dev-F16 with 32gbramlets?
>>101924227FP16 should fit in 24GB just fine
>>101924365I have a 24gb card and I OOM if I don't close google chrome
>>101924365>>101924388please give a secure link to download more ram
>>101924439THE MORE YOU BUYTHE MORE YOU SAVE
>>101924443that is like soooo truuue
>>101924388use your igpu for the desktop
>>101924519How do i do that? is it just changing the display port to the one in the motherbored?
>>101924568yeah, VRAM utilization should show only a couple of megabytes after that
vi
ComfyUI gguf lora yet?
>>101923822realism lora from XLabs
>>101921182are you inpainting the text? having a hard time generating readable text of that size.
strange thing happened to me in these threads,I started hoarding pictures I like. As if will not see them againProbably will never have time to open them, weird.
>>101924817yes, a couple of the initial gens were pretty close to that though. but yeah don't expect 5% of the area of your image to be able to sustain good text or hands etc, which i'm sure you know hence the question
Is not bad bad nudes, but the model still need better fine tunes, I tried Asuka nsfw cosplayer and is good.https://files.catbox.moe/8fyte0.png
>>101925108mama mia thats a spicie meatball
>>101925108go away Teebs
>>101925108she looks a bit young, please don't post that at least without a disclaimer on the catbox. cheers.
>>101925538Asuka literraly has 14 years... Your asuka is too old.
>>101925571mine is explicitly not asuka since she 100% looks like an asian girl wearing a wig, not a half german girl; she is an older girl merely cosplaying asuka. and this must be respected as something with its own appeal. But also i live in a country where i could get fined or worse for accessing this site if someone posts clearly underage simulated photographic stuff :(
>>101923164>ESL
>>101925986at least I'm not a fat bold democrat
>>101926012you're not even american
