Discussion of free and open source text-to-image modelsPrevious /ldg/ bread : >>101671236>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>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>Fluxhttps://huggingface.co/spaces/black-forest-labs/FLUX.1-schnellhttps://comfyanonymous.github.io/ComfyUI_examples/flux>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
blessed thread of frenship
>>101674851>Previous thread lasted 4 hours Nothing like a new model to get anon going
the hired saas shill worked, ahhh /ldg/ is playing 4d chess with itself
Flux bad. Come at me.
>>101675033no need, i bring a solution
Even on fp8 DiT, the pictures are impressive as fuck, too bad I still hate comfyUI and this shit unloads and reload models everytime I make a fucking gen, reeeeeeeeehttps://github.com/comfyanonymous/ComfyUI/issues/2046
all the parameters in the world won't save you from a slopped dataset
>sd3 meme gens>flux meme gens One can foretell the future by looking into the past
>>101675108There's only memes and porn.
>>101675106in flux we trust
Synthetic datasets are the future. The reason why Sora turned out so good, was the use of computer generated data to train spatial and temoral awareness
>>101675134kingdom hearts was made before ai
my eyes gonna break soon
>>101675149picrel me
Have anyone tried running flux on an APU with DDR5 ram?
thread challenge: make the biggest pair of anime boobs you can with flux, bonus points for making me cum
>>101675162missed the button apparently
why does it sometimes make output very blurry?
>>101675181only small booba unfortunately :(
>>101675201
>>101675193yeah I noticed too, sometimes it gets really blurry, dunno if that's a vae issue or something else
>>101675106>just a billion more images >t. eurofag
8gb vram 32gb ramis there any hope for me to run flux?
>>101675181One boobie coming up
>an anime picture of hatsune miku hanging herself, low angle, suicideanyone figured out how to gen people hanging themselves with flux?
>>101675260not a chance lol, even on fp8 it asks for 13gb of vram
>>101675235It's a CFG issue, these models don't work with CFG. Set it to 1.0 if using the regular sampler node.>>101675260If you use the fp8 option.
>>101675278be more descriptive, like "a rope is surrounding her neck" or some shit I guess?
>>101675293>It's a CFG issue, these models don't work with CFG. Set it to 1.0 if using the regular sampler node.Holy shit, I've put CFG = 1.0 on my latest gen and now the speed is like 2 times faster, wtf
>>101675306>an anime picture of hatsune miku hanging herself, a rope is surrounding her neck, low angle, suicideit's getting there...
>>101675335T5 is very literal, try adding rope collar
>>101675340>T5 is very literal, try adding rope collarcan we replace T5 with something else or?
>>101675316
Can't believe we managed to get a dalle3 with better quality pictures at home, and it's only using 12gb of vram on fp8, feelsgoodman
>>101675340>T5 is very literal, try adding rope collari didn't know that, ty for the tip>an anime picture of hatsune miku hanging herself, a rope is surrounding her neck, she is wearing a rope collar, low angle
>new model>24GB VRAM>don't feel like swapping the 7900XT in the linuxbox with the 7900XTX in the gaymingbox
>>101675328That's what happens when the model only needs to do half the work because there's no negative prompt.
what's going on with audio generation? I remember a model coming out months ago and then I suddenly stopped hearing about it. where are the threads? did (((they))) shut it down?
is flux 16ch?
>>101675389https://comfyanonymous.github.io/ComfyUI_examples/audio/Stable audio is a thing that exists.
>>101675388what? if this model doesn't work with cfg, that means it doesn't work with negative prompt?
>>101675401Good question
I tried to change to 1400x1400 resolution and I got this error
>>1016754221536x1536 worked for me. but the image looked very overcooked
>>101675408Exactly. Positive prompts only.>>101675422Yeah some specific resolutions won't work at the moment, I have to fix that.
bigma.......
>>101675441>Exactly. Positive prompts only.that's insane... negative prompts are really important though, and flux can't do that? damn...
>>101675453Flux only makes Bigma stronger because it's motivation. AI models are an arms race.
>>101675441i want to kill that creature
>>101675306>>101675335You hang yourself with a "noose" fucking ESL faggots.
>>101675472>AI models are an arms racenope, it's a finger race now. both kolors and flux have set a standard for hands
>>101674851
I guess artists have no reason to complain if the models can't generate a painting!
Have we finally found the new base model friends?
>>10167547She can defend herself.
>>101675520The issue is how expensive it will be to finetune. Joe smoe cant get it done on his 3090 anymore.
>>101675293>It's a CFG issue, these models don't work with CFG.the API has a CFG setting thoughhttps://replicate.com/black-forest-labs/flux-dev
>>101675520don't know, gotta wait 2 more months and see what happens
>>101675543dev version works with 2 ish cfg. its not that it does not support it, its just that distilling requires lower cft just like lightning models
>>1016754361536x1536 looks fine for me on schnell, but like a blurry mess on dev
>>101675501not sure if lying, schizo or incompetent
>>101675556but I'm running flex-dev, it's also a distilled model?
>>101675520>>101675542The base model already requires 24GB of VRAM.
>>101675542QLoRA for diffusion image models when?
>>101675558same on dev .. not sure whats better?
>>101675558maybe its scheduler related? ddim_uniform looked more cooked than picrel with simple scheduler. flux dev fp8
>>101675580Can run on 12GB vram with 8 bit.
So Flux doesn't support negative prompts? The example Comfy workflow has a positive prompt only.That's a pretty huge limitation even without taking into the account the slopped aesthetic and inability to do painting.
>>101675541https://www.reddit.com/r/comfyui/comments/17h66ld/comment/k6mxxac/?utm_source=share&utm_medium=web2x&context=3Can you consider making multi-gpu possible now? we have a great model that is tough to run with only a 24gb vram card, would be cool if we could use multiple cards for that model and not rely on fp8 instead
guess i'll just shleep till bigma then
>>101675616It does, guy did not know what he was talking about. CFG just needs to be low on distilled models
>>101675608>>101675558so only dev makes blurry pictures like that? dunno if I should switch to schnell? what's the difference between the 2?
>>101675500give her armpit hair
>>101675616If you have a prompt you can have a negative prompt. It's just math: positive - negative = prompt
>>101675610could be karras on dev seems to not work at all for me .. also is it a different data set? and what res does the data set have? details can pop 1536x1536 but it hugely depends on the subject you prompt
>>101675641schnell seems to be the sdxl lightning equivalent
>22.28s/itThis is painful.
>>101675543That's not real "cfg" it's an extra condition passed to the model. https://github.com/comfyanonymous/ComfyUI/commit/a531001cc772305364a319a760fcd5034e28411aYou can use this new node to play around with it.>>101675620We'll see.
>>101675641omg... change cfg. schnell is like using a lightning model, it needs lower cfg
>5g vramIs it over for me?
https://blackforestlabs.ai/announcing-black-forest-labs/>FLUX.1 [dev]: FLUX.1 [dev] is an open-weight, guidance-distilled model for non-commercial applications.>FLUX.1 [schnell]: our fastest model is tailored for local development and personal use. FLUX.1 [schnell] is openly available under an Apache2.0 license.Uh oh...
>>101675680yup :D
>>101675558transformer-based models just seem way worse at out of distribution resolutions in general than diffusion models, I found the same thing with SD3like when you make a diffusion model do an OOD resolution the texture of the image will still be fine, it'll just loop on the composition, so if you can use tactics to prevent the looping then you can push the limits of resolution quite wellbut when you give a transformer model an OOD resolution it just starts falling apart, it doesn't loop, the image just turns to incoherent pixel mud
>>101675678>omg... change cfg. schnell is like using a lightning model, it needs lower cfgboth need low cfg, flex is also a distilled model
>>101675696damit was nice to know you ldgbros
>>101675706that's cute but in practice it's stupid
Show me Flux's "double exposure"
>>101675671>>https://github.com/comfyanonymous/ComfyUI/commit/a531001cc772305364a319a760fcd5034e28411a>You can use this new node to play around with it.I have no idea what it is? how do I make it work?
>>101675727you have not articulated yourself clearly
>>101675641damned if id knew, the huggingface page only lists very little details, but the same prompt can have very different results on both models, dev produces anime on one prompt for me, while schnell makes photorealism ... schnell works on karras, then dev fails on karras, I wish they had some documentation
>>101675731seems fine
realistically can bigma even reach this level? if it does i might die of hype.
>>101675743there's this >>101675685
>>101675749I'll count that
>>101675542I think normies can do smaller finetunes with less than 50 bucks when they rent H100 or two for a day.
>>101675749>>101675763Not as good as base sigma desu but better than the others
>>101675767numb brain here, can you finetune a distilled model?
>>101675831That's a not cherry picked lazy gen. It's the first thing that came out.
>>101675839yes
>>101675845Relax anon we are on the same side
>>101675858My point is you can't really qualify how good it is because it's one sample. It knows the effect, I'm sure you can achieve something better.
>>101675671https://github.com/comfyanonymous/ComfyUI/commit/a531001cc772305364a319a760fcd5034e28411a>You can use this new node to play around with it.First time in my life I use ComfyUI, how do you load this node?
>>101675750The 1.3B model will not be able to reach the level of a larger model. Its primary advantage will be that it's truly uncensored and will have a lot of pop culture knowledge. Who knows what the official Pixart model will target, I assume they're going to do a 24 GB squeezer too and it might be really good, they certainly proved they're kings of efficiency.
>>101675876>CLIPTextEncodeFluxdouble click on a blank spot and try searching for that
>>101675876>>101675734You update, double click on the canvas and search for Flux and then click on CLIPTextEncodeFlux.Then you replace the CLIPTextEncode node in your Flux workflow with it.
if the official sigma successor is a 24gb squeezer what will we be calling it?
>>101675922Squeezema
Both Flux models seem unable to generate illustrations of flat-chested lolis in a swimsuit.
Question: If someone makes a finetune for flux-schnell, can someone use it with Dev?
>>101675938Yeah, but they're able to generate flat-chested children in a swimsuit.
i jus want a bigger pixart model grumble grumble
>>101675938FBI please
im using a realistic pony checkpoint, and i want more vivid colors so i dont have to adjust it in photoshop. should i be trying to find a color vae, or a lora, or just try to force it with prompting/
>>101675919>>101675920like this? so the "guidance" thing is like the cfg of distilled models?
when i get my hands on flux itll be over for you
ugh these models that were trained on AI captions are doa for anyone who wants to generate paintingsauraflow had this problem and flux seems to as well, where they don't even know the names or styles of famous dead public domain painters
>>101675968SDXL is fundamentally fucked when it comes to colors
>>101675954They really are. Only problem is, I don't care about realistic photos.
>>101675959Buy me a computer with four H100s
A cat riding a bicycle with wheels made of pizza on the surface of the Moon. Planet Earth is exploding in the background.
>>101675974sorry pal, show's over, everybody's gone back to waiting for the next thing mode.
>>101675980at this point, more pretraining must be done on those models to add the styles and stuff
>>101676008what about a few words of encouragement instead?
>>101675989that sucks, are there any techniques people use to adjust? like control nets or something? i cant go back to 1.5 after seeing how sdxl can generate feet and stuff
>A sculture of Donald Trump made only with bottlesWhat the fuck??
>>101675970Set CFG to 1.0 or use the example workflow.
will we be getting native pixart support on comfyui one day?
>>101676055you're prompting on clip_l anon, that's not the good one, go for the t5xxl slot
>>101675671>>101675920you have up mixed t5 and clip, they have an opposite effect
>>101675963Or flat-chested anime girls in a swimsuit / bikini / underwear in general. Just try, Flux will always add fairly large boobs. How do you control breast size with these models?
>>101676064Does the dual clip loader run t5 on CPU/RAM by default? That's what I want and it's what seems to be happening, just want to confirm
>>101675671Is there a way to remove the "clip_l" slot with a "remove button" or something? that shit is useless and take some space at the end
>people using flux pro in replicate and other API sites>don't realize that schnell is what will get any support from finetune community, if it even get it, since it's still very unclear how hard it is to train
>>101676161just like sd3 huh
>>101676161what about flux?
>>101676168cept with apache 2.0 license and not being shit
because flux is distilled, it means it's supposed to work at 4steps?
>>1016762248
desu I don't see much difference between pro and flux, It's rare that companies are making a local model at the same level as their "best" API model
>>101676257how do you know it's 8?
>>101676326default workflow had 8. Should work at 4 as well just with less detail
>>101676316neither of them is remotely in the style of picasso like you asked for, thoughnone of these new AI-captioned models coming out seem to understand style for shitold models were great at style but had shit coherence, now it seems we've swapped the two
>>101676055Negative prompt still doesn't seem to work with this setup. What do
>>101676316You mean "dev" and "pro"?Also, try to remember that "dev" is not Apache 2.0 and will most likely never get big costly finetunes like Pony, because it will costs like hundred thousand dollars and you can't monetize it in any way.There is third one, which is "schnell", and that is the truly open version, but it's distill, few-step model that requires at least 16 GB of VRAM to operate and will be a bitch to train, and when you use schnell, you realize that training is what it needs. It lacks styles and concepts.
>>101676345Sounds like steps = 10 is the sweet spot
>>101676360Artists stuff are removed due to legal concerns. Its still a very grey area.
>>101676372>There is third one, which is "schnell", and that is the truly open version, but it's distillFlux is also a distilled model https://blackforestlabs.ai/announcing-black-forest-labs/>FLUX.1 [dev]: FLUX.1 [dev] is an open-weight, guidance-distilled model for non-commercial applications. Directly distilled from FLUX.1 [pro]
>>101676382No it's because of AI captioningThat's why they don't know dead public domain artists from the 18th century either
when I change the weight_dtype between fp8_e4m3fn and fp8_e5m2 I see no difference, what's it supposed to do again?
>>101676402yeah but it knows celebrities though so...
>>101676402>not trained on data>REEE its the ai captioning. SLOP SLOP SLOP
>>101676407show us 2 pictures with the exact same seed and settings to be sure there's no difference
>>101676419meds
>>101676402there is no way Picasso fails to be picked up by even v1 of llava. they were removed.
>>101676449If that's true, then not coding in an exception for dead public domain painters is lazy as fuckit's a few hundred names at most, assembling a list to make a whitelist of allowed artist names would have taken 10 minutes
>>101676392They are all "flux". There is three flux.1 variants.>Pro>Dev>SchnellPro is what you will never get to run locally. It is the one they are monetizing. Dev is what you *can* play with locally, but you can't monetize and you agree their license (https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/LICENSE.md) and it restricts you. Schnell is the one you can do what ever you want with.
>Requested to load FluxClipModel_Is this mendatory to load clip_l? Why does this model has 2 text encoders?
>>101676486>lazy as fuckthis has been the case for local datasets since sd 1.4, can't really expect much to change.
>>101676490yeah my b, what I wanted to say is that dev is as much as a distilled model as schnell
>>101676316what would be the equivalent of interval in sd language, a loopback? a detailer node?
>>101676490Pro is also the base model. I don't understand why more people aren't concerned about that.
>try to make Kolors draw closed lips challenge (impossible)REEEEEEEEEEEEEEE
how are you supposed to get rid of the blurry ass backgrounds with no negative prompt
>>101676538yeah, it's retarded to pretend that negative prompt isn't important, it fucking is
>>101676538>how are you supposed to get rid of the blurry ass backgrounds with no negative promptyou put a low cfg value like 1.5 so that you can use negative prompts
>>101676547ugh, you must be a dallejourney shill
>>101676570I'm not lol, I fucking love this model, we literally have dalle3 with better realistic humans at home, it's just that negative prompt is really important though
>>101676583https://github.com/comfyanonymous/ComfyUI/commit/a531001cc772305364a319a760fcd5034e28411a
>>101676593that doesn't change anything, you can't use negative prompt at cfg = 1, regardless of the guidance scale
>>101676602put cfg at 1.1 then
>>101676602This, this model needs really low cfg, like 1.5 or a bit lower>>101676612
>>101676642>>101676612yeah, even at 1.5 it starts to hurt the output
>both of the released sets of weights, the Schnell and the Dev model, are distilled from the Pro model, and probably not directly tunable in the traditional senseIt's over
>>101676724you can tune distilled models, just takes longer to make up for the loss
euler_a doesn't seem to work on flux-dev
>>101676766tried any custom SCHEDULERS?
>>101676772no lol
wonder if an ipadapter could fix flux not being able to do painting stylesyou could use ipadapter to do style transfer maybe with a painting as input
>>101676766ancestral samplers don't seem to work with SD3 either, maybe there's something about them that makes them only work for diffusion models and not transformers?
I really don't recommand to use the fp8 version of the text encoder, this shit is ass
>>101676851it's been a while I haven't used SD models, and there's a shit ton of samplers now, is there one that is above the rest or do they all have the same level?
What s/it we suffering at boys
>>101676752Did they even release training code?
>>101676851Ancestral samplers work on sigma
>>101676852So we have no other choice but to use two text encoders for this model? can this be two t5xxl instead of t5xxl + clip_l?
>>101676888I thought pixart was a diffusion model
The advantage of having dalle3 at home is that I won't have to be cautious of triggering any cucked censorship, feelsgoodman
First time ever using Comfy, how the FUCK do I load Flux properly? All the pics it's been spitting out so far look dark and noisy as fuck.
>>101675377>>101676953This is extremely dangerous to our democracy.
>>101676975download this picture >>101676953and click on load on ComfyUi, it will use the working node and you're good to go
>>1016769964chinz removes pic metadata, got a catbox?
they obviously did not fully strip artist namesprompt here was "art by Makoto Shinkai", you can see it clearly recognizes himso I think the anon that said the lack of artist knowledge was due to AI captioning rather than intentional stripping is likely correct
>>101677006oh my b, there you go anonhttps://files.catbox.moe/k50f9f.png
sometimes it add some artist signatures, this one doesn't seem to exist though
>>101677028Thanks, I'll give that a try.
>>101677014i think the answer is both in a way. the autocaptioner loses some niche ones while the pruning is hitting bigger ones. i've seen at least 3 different people try picasso and fail each time. i mean if it can't do that then idk what to say, maybe it's an architectural issue or something i don't really know.
>>101677014another test, this was>art by Margaret Olleyagain it clearly knows this name, because this is the kind of thing she painted, still life of flower vases and suchso it's not total doom for painting enjoyers
>>101677059Ahh man, nice texture on this one too. Oil paint with a canvas texture. So it can do that after all.
No fucking way... it doesn't know who Rei Ayanami is...
>>101677059>it clearly knows this name, because this is the kind of thing she painted, still life of flower vases and suchI tried using that style and it didn't work, it didn't include Trump too but only his hair onto Hatsune Miku top kek
>>101677059Yeah that looks very good, yet the ones i've tried for obvious artists have come out terrible. I wonder if this is similar to that Auraflow cat thing. Where 1/4 of the synthetic dataset produced a 'no nsfw cat' so it wound up reflecting in random prompts due to it being assigned captions regardless. Perhaps there was synthetic data used from garbage midjourney gens where people throw in "By Yoshitaka Amano, by Pablo Picasso, by Wlop, by Greg Rutkowski" and out comes a slopped-up 1girl that looks nothing like any of them. If they trained on those captions directly, then it would make sense that the more popular artists have their styles completely watered down from misuse. This is now my leading schizotheory after thinking about it more.
>>101677014interesting
I'm late. Does it do NSFW? How pozzed is their dataset? What about overall anatomy.
>>101677153>Does it do NSFW?Yes> How pozzed is their dataset?Doesn't know much style and anime characters :(>What about overall anatomy.Really great, API level great
>>101677014>hey obviously did not fully strip artist names>prompt here was "art by Makoto Shinkai", you can see it clearly recognizes himit feels like as long as you put a character that was known for its own style, it's hard to make it translate to a different style, that sucks
>>101677137so you're saying my esoteric knowledge of classical oil painter names that normal people have never heard of is finally going to be useful
>>101677141I swear to go you give a random guy this picture it will believe it's from a random anime, that's great
>>101677137that's an interesting theory anon, another reason to not use synthetic data, it will train the model on wrong representations of real artists
>>101677188I wish but I tried that on Winterhalter, Albert Lynch, Hans Gude, etc and it just failed every one. It doesn't even do them wrong, it just doesn't do them at all
How do we finetune it?
>>101675377>>101675615Works even on 10gb 3080 with 8bit :)(comfy launched with --novram probably helps too)
localchads won
>>101677280What speed are you getting? 2 hours per image?
is there really a need for a diffusion general on almost every board, even /b/? seems suspicious to me, it's very under powered and not really "ai" but it's being shoved in everywhere on 4chanevery single one of these generals is bumped by bots
>>101677188>>101677212>Painting of Super Mario's head, classical painting by Hans Gude, Winterhalter, Albert Lynch, Michelangelo. I either get this generic 3d render on top of a paint-ish background, or some random airbrushed guy that looks nothing like mario on top of a paint-ish background. The model being unable to put characters into a 'le classic painting' style is really odd given this was one of the very first uses of SD back in 1.4 and WD 1.2 with the victorian booba girls.I really don't think it's 'style pruning' here as even with every artist name pruned the autotagger should recognize 'a classical painting'. Am I a promptlet? Is such a prompt not enough?
>>101677299The /b/ one is definitely needed
>>101673628>>101674725If synthetic images make you that upset, why not train your own models then rather than getting angry at others?Maybe you can convince someone gullible to throw you some training money as you're obviously an expert on the subject.Meanwhile others are going to keep using real images and synthetic images in our models..
https://blog.fal.ai/auraflow/I'm testing if flux dev is as good as auraflow at promt understanding, left is flux dev, right is auraflow>A photo of a beautiful woman wearing a green dress. Next to her there are three separate boxes. The Box on the Right is filled with lemons. The box in the Middle has two kittens in it. The Box on the Left is filled with pink rubber balls. In the background there is a potted houseplant next to a Grand Piano."
>>101677314no need to seethe over being called out. just improve next time, it's okay to be wrong
>>101677307Auraflow has this exact problem too, where content overwhelms style
>>101677307Cascade, which seems to do a better job at it. Dataset, architecture, both?
>>101677317>a cat that is half orange tabby and half black, split down the middle. Holding a martini glass with a ball of yarn in it. He has a monocle on his left eye, and a blue top hat, art nouveau stylethat's a shame that it doesn't know much about styles of actual characters trivia, because it follows prompts really well
>>101677356data, always
>>101677356>Dataset, architecture, both?I'd go for synthetic captions, they probably used too much of them and destroyed all the artist names in the process
>>101677192Harder to gacha text when it takes 2 minutes per image is the only issueAdd LoRA support and it opens the road to fake spoilers and gaslighting people about shows kek
>>101677307>sonic the hedgehog. impressionist style painting with wide brush strokes. visible paint daubs, canvas texture
>>101677476about the furthest thing from impressionist style out there, but at least it got the brush strokes
>>101677476>>101677495>impressionist style painting with wide brush strokes. visible paint daubs, canvas texture. homer simpsonyeah, not impressionist at all. but seems to work for getting it to do an actual painting and not digital art or a render
>>101677356>Cascade, which seems to do a better job at it. Dataset, architecture, both?
Have you tried other samplers than euler? Maybe there's one that works great with flow dev
>>101677411Don't think so entirely, as cogvlm can recognize a painting even if it loses the artist. Certainly the case for it not understanding Winterhalter, etc, but not for painting in general. It seems you really need to force-feed it the idea of a painting or else it fails especially when combined with characters. I'm able to do "A detailed classical painting of a variety of fruit" and it comes out pretty fine. Likely something architecturally I guess if it starts to get stubborn when combined with different concepts.
>>101677549I tried dpm2 that someone suggested, and it came out pretty fried and nasty. euler remains eternally winning amidst a sea of snakeoil
>>101677575works on my machine
flux is SEXO
>>101677630it almost always have perfect hands, that's so cool to see, looks like hands are a problem from the past now
>>101677515kek
>>101677651lmao
>>101677587>works on my machineBut does it work on your atachine anon?
>>10167729314s/it so around 4-5 mins per image at 1024x1024, not great but seems about normal
>>101677288localchads always win
>>101677683Huh, way better than I thought.
That's insane, it reproduced the roman number of the clock perfectly
fp16 is too big but fp8 is too different to fp16, how about fp12? is that even possible?
Is there a quick setup guide for flux on comfyui? What folders do I put each file in, what settings to use.
locals walks so saas can run. as a midjourney and dall-e user, i thank you all for your service. none of our amazing closed models would be possible without your free research.
>>101677751yeshttps://comfyanonymous.github.io/ComfyUI_examples/flux/
>>101677759I'm crying
>>101677718fwiw this is on Linux and I haven't updated the Nvidia driver in a while so as far as I know there is no VRAM->RAM offload like on Windows. Sitting on 9.8gb VRAM usage though lol
Can Flux generate Laura Kinney in a bikini?I might just stick with pony
>>101675567catbox?
Why did they go for 12b, it's a bit too big for my 24gb vram card ;-;
>>101677828works on my 3090 THOUGH
>>101677811its by far the best we have gotten with a base model ever. We just need one of the big finetuners to take the leap and we will have THE next gen model
>>101677828I'm running it with 20GB VRAM.
>>101677835mine too (3090 chads rise up) but barely. the vae decode takes ages in fp16 because the vram fit is so tight. I have to drop to 8bit if I want my computer to be usable at all while I'm genning.
>>101677811I highly doubt it knows who that is
>>101677840Anons said you need a stupid amount of vram to run it
>>101677835>>101677843for me it reaches the limit and everything slows down, fock, maybe I should disable hardware optimisation on chrome or something
>>101677828It fits, and 8 bit fits on 12GB and the difference in quality is minimal
>>101677851If 12GB is a stupid amount I guess. You could edit some files and try it at 4bit to maybe get it to work on a 6GB card
>>101677851
>>101677854>>101677843>>10167783524gb is still not enough though, it only run the image model, if you want to add the text encoder it would need an insane amount of vram
>>101677873follow the instructions anon... even with FP16 everything it fits in 24GB with room to spare
>>101677881no, only the image model is fitting the vram, try to do a --highvram flag and you'll see your gpu will be oversaturated
>>101677873Yeah I'd imagine we would get to a point where we would need to work on multi-gpu support
>>101677881nta but nah, the t5 model is being run on your cpu/system ram
>>101677849Pony it is
>>101677902Stop doing retarded shit and just use someones workflow from the 1000 catbox links / the site
>>101677904>Yeah I'd imagine we would get to a point where we would need to work on multi-gpu supportI'd be ok for an hybrid cpu/gpu inference on the text encoder, instead of putting the whole text encoder to the cpu when it can't fit tonto the gpu, that's something that has been done on the llm community
>>101677929he's right though -> >>101677926
>>101677873Just run in on your second 3090 that isn't connected to the monitor. Surely you have one just like all of us?
>>101677953>Just run in on your second 3090 that isn't connected to the monitor.how? there's no multi gpu support on comfyui
>>101677960I meant load the model fully on the idle GPU instead of the one that runs your system.
https://huggingface.co/motexture/FluxDiffhttps://github.com/motexture/FluxDiffhttps://blackforestlabs.ai/up-next/not seeing any posts about this
>>101676064honestly wishing I had negs right now because it keeps giving me heart pasties. hydit fennec ears > flux fennec ears. these ones aren't nearly as fluffyhttps://files.catbox.moe/s6xb8y.png
It's weird they decided to use t5xxl as the text encoder, this shit is 2 years old at this point
>>101677984those guys will revolutionize this ecosystem, they're insane, if it has the same quality as their image model we'll be eating so good... our true savior
>>101677985>honestly wishing I had negs right now because it keeps giving me heart pasties.try to go for a low cfg like 1.5 so that you can access to the negs
>>101677984oh no no no no no sorasissies...
the skeleton of voice synth has all but turned to dust....
is it worth paypigging for runpod?
>>101677984HOLY FUCKING KINO
>>101677984if the text to image gen needs 24gb to be good, there's no way we'll be able to run high quality text to video models, unless Nvdia decides to give us more VRAM... LOOOOOOL
goddam I wrote "adult" on the fucking prompt, why is it showing me a kid?
Which one is better? fp8 e4 or e5?
>>101678140this model has a big problem with blur, that's why a negative prompt is important in the first place...
We are so backing back it unreal . Open source bros we won
>>101678140>>101678183the problem is your scheduler, select "simple" and it will work, if the number of steps are too little I think it also outputs blurry images
https://huggingface.co/Kijai/flux-fp8/tree/mainflux dev in fp8, 11.2GB
>>101678167"We also see that FP8-E5 is never the best format for inference; even for the transformer layers with significant outliers, the FP8-E4 format is better"https://arxiv.org/pdf/2303.17951I'm not sure how applicable this is for us though, that being said, the VAE uses e4 so thats already 2 data points supporting e4
>>101678228oh shit
>>101678199I went for 20 steps and it works fine now>ohh it's a distilled model it should work perfectly at 8 stepsthat's a fucking lie, 20 steps seems to be the sweet spot kek
>>101678245The DEV model isn't distilled for low step count, only the other one, the DEV one peaks at 40~50 steps according to the developers
>>101678067Get multiple 3090s. /lmg/fags are hacking together 4x3090 systems so they can use 100b+ local models. I'm actually super excited we finally have a larger diffusion model. These 1-2b models were never gonna cut it. None of the inference backends support multi-GPU yet, but conceptually it's not hard, you literally just put half the layers on one GPU and half on the other, and use pytorch hooks or whatever the fuck to make sure the activations automatically route between them. Actually HF Diffusers might already do GPU splitting, certainly every single HF Transformers model can do it out of the box.
>>101678260>the DEV one peaks at 40~50 steps according to the developersohh... I should stop going into the 20 steps then, can you give me the source of that though? that looks interesting
>>101678228BTW anons. If you already have the original weights, you get the same outputs as this by using the --fp8_e4m3fn-unet flag.Also, if you plan to download and use that, also run it with --fp8_e4m3fn-unet, or else it'll try to take up fp16 amounts of space for some reason.
>>101678228what's the difference with downloading the whole model and putting the fp8-e4 weights though?
>>101678252neat
>>101678269>Get multiple 3090s. /lmg/fags are hacking together 4x3090 systems so they can use 100b+ local models.I already a 3090 + 3060, and unfortunately I can't use the 3060 because there's no multigpu support on comfyUI
>>101678269yeah, I'm surprised it hasn't been done before, the llm community can do that for a year now
>>101678274I thought I saw it in the HF model card but it isn't there... still, I distinctly remember reading that, but as you can see its still very good below that
>>101678315kek
>>101678291You wont have to read 24gb of garbage, and if have 32gb of vram this also prevents having to swap to disk while loading for the first time
>>101678377oh nice, I still have to specify the e4 weights or should I put "default" for that one?
>>101678315Same setup here. Poor little 3060.
>>101678415It shouldn't matter, but I don't know, I tested and leaving it at e4 matches the original model
>>101678315the t5 loader from ComfyUI-ExtraModels has the option of loading it to a secondary GPU, and it works well. There's no way to use it with Flux unfortunately (node output is the wrong type). But it proves it's quite possible and Comfyanon just doesn't want to implement it.
Fresh bred when ready >>101678250>>101678250>>101678250
>>101678437fuck...
