Discussion of free and open source text-to-image modelsPrevious /ldg/ bread : >>101930742>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
bros why the FUCK is comfy sudddenly so slow when using controlnets?the first time i click "queue prompt" it gives me an OOM error, then when i click it again it goes through but it's slow as fuckit wasn't like this a couple of days ago, no OOM errors or anything - i'm using the exact same workflow now, even reverted back to the commit it was on (problem happens on master too though), only difference is that today i'm getting 80s/it and it's absolutely UNUSABLElooks like the controlnet loading is making it hit the VRAM limit but then why did it work fine a few days ago and how do i fix it?i'm on a 3090 if that mattersi really don't want to resort to using flux quants because of the loss in quality, and no i don't care how "barely perceptible" it may seem, it's still there and i don't want it messing up my highly specific prompts
>mfwwhat’s ldg’s current meta for guidance/cfg with gguf?
this bread is fucking COLD
4x-AnimeSharp test
someone bake a real bread please
>>101933406>but then why did it work fine a few days ago and how do i fix it?Blame Windows?I think that if you OOM there's an nvidia driver setting that makes things slow. So something changed that uses up more VRAM in general use.
flux1-dev-fp8.safetensors works pretty well for an all in one file
>>101933488Go for Q8_0 instead, it has the same size and is closer to fp16 in accuracy
>>101933504Q8_0 oom on my machine but FP8 work
Now I've become GuidanceNegative, the destroyer of slop
>>101933485>something changednot very insightful :\>>101933488try out gguf>>101933517you're combining it with dynamic thresholding?
>>101933513that's weird, they're roughly the same size, did you put the text encoder into your second gpu or your cpu?
>>101933504Is there a comparison like this but with Q4_1 and Q5_1 added?
>>101933525I get this error when I try to use the q8 file: You do not have CLIP state dict!do I have to add the clip files for this model? other was all in one
>>101933525>you're combining it with dynamic thresholding?No, with Tonemap
>>101933536>forge errorno sir I cannot help. I'm a comfy boi>>101933542>with Tonemapwut
>>101933556>wut>he doesnt know
>>101933532there will also be the QK quants, this image will be to long kek, if you want to know which one is better than the others, just look at some test made by the /lmg/ fags, the least % on the last column, the better >>101933556>wutTonemap is an alternative to DynamicThreshold, it burns the image way less https://reddit.com/r/StableDiffusion/comments/1estj69/remove_the_blur_on_photos_with_tonemap_an/
>>101933556I use comfy too but I am just testing both, is q8 better or worse than full fp16 in comfy but weights set to fp8?
>>101933530I have this, I tried --cpu-vae
>>101933530Maybe FP8 gets to do math as FP8? Q8_0 converts a tensor at a time, which uses a bit of VRAM.
>>101933569>https://reddit.com/r/StableDiffusion/comments/1estj69/remove_the_blur_on_photos_with_tonemap_an/thanks kind stranger!
>>101933517Why does everyone use Anime to show off their negative prompt hacks? Come back to me when it works with photos and doesn't overburn the absolute shit outta them
>>101933577no, not like that, use some nodes to force each one of them (unet, clip, vae) to a specific part of your machine (gpu or cpu)https://reddit.com/r/StableDiffusion/comments/1el79h3/flux_can_be_run_on_a_multigpu_configuration/
>>101933504G8_0 is bad at human hands and details
>>101933581>Q8_0 converts a tensor at a time, which uses a bit of VRAM.no, it's more because Q8_0 is a 8.5bpw (bits per weight), wheras fp8 is a 8bpw, Q8_0 is a little bigger in consequence >>101933504
>>101933586>Come back to me when it works with photos and doesn't overburn the absolute shit outta themthis? >>101933569 it removes the blur on photos https://imgsli.com/Mjg3OTYw
>>101933596it's flux's fault, fp16 also has horrendous hands in this example
>>101933616I've tried your tonemap hack, all my photos turn red and horribly overburned even with all the updated scripts
character creation slider when?next year? or 2030?I mean 2d_to_3d and 3d_to_2d is the future
>>101933628show me a screen to your workflow with such burned pictures?
>>101933596>>101933627>bad handsI never had any problems. just prompt better lol
Pixelart? How good are they? Especially lows res sprites and tiles like snes era
Does low vram mode affect quality?
>>101933649what are you doing here anyway? go back to your thread
>>101933670nope, every output is the same and deterministic
>>101933660what model are you asking about?
>>101933588I'm ooming sir
>>101933660I'd say pretty good
>>101933686that's so weird, how much vram do you have?
>>101933675God has no nation. all domains are God's domain.
>>101933504>nf4 is the only one that didn't put a stupid blurred background in a DRAWING of all things.That one for me.
>>10193371112
>>101933720that's weird that you can run the fp8 with only 12gb of vram, it asks for more usually, maybe its loader is putting the surplus on your ram wheras the GGUF loader doesn't
Forge support fp4 but where are the models?
>>101933718kek>>101933737nf4 sucks, go for Q4_0 instead (it's the same size but better quality) >>101933504
>>101933686Why are you trying to force the actual model onto cuda:0?
>>101933753because he wants to run the model into his gpu? usually that's what you should do, put the text encoder in the cpu and the unet/DiT model in your gpu
>>101933743FP4 not NF4
>>101933731idk>>101933753Because I don't want super slow gen?
>>101933768Q4_0 > nf4 > fp4
>>101933743fp4 (floating-point 4) is not the same as nf4 (normalized float 4)
>>101933766>>101933771That stops it from ever being unloaded while also disabling lowvram though. You should never pin anything to cuda:0 and I regret ever adding that as a possibility lmao.
>>101933791>You should never pin anything to cuda:0 and I regret ever adding that as a possibility lmao.you're the one who made that script anon?
>>101933783I know, but I want to see it.Also there's int4 and int8 models on HF, idk how to use it.
>>101933789the point still stands, nf4 sucks compared to Q4_0, and fp4 is worse than nf4, so fp4 sucks ass even more
>>101933743>nf4 proof?the person who is testing it obviously used comfy therefore is not a valid comparison that's why he gets a different result, also the prompts in forge get different results compared to comfy
some league gens
>>101933817that's a fair point, the same prompt comparison should be used for forge I guess
>>101933800Yes, currently trying to get LoRAs to work if pytorch would stop stripping my custom bullshit tensor metadata at every chance it gets.
>>101933837oh, that's why I OOM when loading loras, that's because of your script? :(
>>101933791>I regret ever adding that as a possibility lmao.don't regret anything anon, as a vramlet I managed to use bigger quants because comfy's automatic memory management sucks ass and your method made it feasible for me
>>101933840>>101933850To be fair, "set MODEL device" was a fork by some other guy that also tried to offload the unet - something which I removed in the final version since it kept breaking kek.That script was only ever meant to put either the T5/CLIP models or the VAE on a second GPU or the CPU.There is also this which supposedly works better, though I haven't tried it: https://github.com/neuratech-ai/ComfyUI-MultiGPU
What's interesting is the gguf implementations are python versions. We could get even more speedup from pure C/C++ implementations of it, as it was originally intended.
>>101933872c++ inference is a myth. its not real
I have a comfy lora workflow but the realism lora seems to be doing nothing. is there a proper workflow that is recommended or what.
>>101933871>There is also this which supposedly works better, though I haven't tried it: https://github.com/neuratech-ai/ComfyUI-MultiGPUI'm not a big fan of that one because it doesn't adapt, for example the nf4 and gguf loaders won't work wheras your force/device will work for all loaders
>>101933876aww she's so cute
>>101933884>c++ inference is a mythhttps://github.com/leejet/stable-diffusion.cpp
>thread schizo infestationoh no
>>101933913looks promising. comfy should copypaste this
>>101933899pixel art + anime or chibi style works really good in flux. in sdxl/pony you have to use the pixelization extras feature to get good pixels.
>>101933958>pixel art + anime or chibi style works really good in flux.on the other hand, flux is too biased to put chibi style everywhere, even on drawings that have the supposed "50's comic book" style to it >>101933504
>>101933953He should replicate the command line mode, yes.
>>101933971you can get good stuff if you specify a specific style or type of comic book.
>>101933982care to show some prompt examples?
>>101933982I think that one was just Miku Hatsune in the style of a 40s Marvel comic book, it wasn't that complicated: specifying the time period gives you a much different aesthetic.
>>101933988ok pngchunk sorta works with flux files:Miku Hatsune appears on a 1950 comic book. Miku is in a green ballroom dress, holding up some carrots. Text at the top of the comic says \"Miku Hatsune\" in bold text. The top right part of the image has a Marvel comics logo. She is standing in front of a building in New York.
what absolute retard thought it would be a good idea to randomise the NEXT seed when queuing a prompt in comfyui?i've been fishing for a good gen for over half an hourfinally got one, but saw in the preview that it could use some tweakingdidn't want to wait 10h for the gen to actually complete so i cancelled it, fixed the seed, tweaked it and went to regen, but the result was completely differentafter testing it seems that clicking queue prompt immediately fills the seed with something it'll use on the NEXT genwhose retarded idea was this?
>>101934004that prompt is too "easy" for flux because the style you specify also aligns with the rest (Marvel comics logo, Building in New York), if you try a different setting like Idk in Paris and you don't help it further by not specifying the Marvel logo it will collapse
>>101934017You decided that. There's an option in the settings to choose how it behaves.
You don't have to reply to the schizo just laugh at himRead the OP for why
I'm... I'm gonna... I'm gonna OOOOOOOOOOOOOOOOOOOM
when i'm using DPM samplers in comfy, should i be adding a scheduler? DPMPP and DPMPP3 seem to be completely unusable, trying to recreate some gens from forge and i can't pull it off 100% because it deep fries the fuck out of the result.
>>101934017yeah it is 100% retarded
speaking of samplers/schedulers, is there a good comparison chart between them anywhere?i only remember back when a1111 had euler and dpm (up to 2?), the rest are all new to mewhat's generally best?
>>101934017can't agree more with that, that's why I always put on "fixed" and if I'm not happy with my gen I increase the seed by 1 kek
>>101934088there's this if you're interested https://reddit.com/r/StableDiffusion/comments/1em7zy3/testing_the_samplers_schedulers_with_a_xy_plot/
>>101933504I kinda feel these sorts of comparisons are pointless, looks like the same as comparing samplers, because you can squint at every fine detail and say "a-HA! choice X is clearly better than choices Y and Z!" and then do a different prompt with a different lora and it's entirely the opposite.it's just like, I don't see how you can make any persistent inferences one way or the other, like "if you use this choice you're gonna get images that do this", it's all just irreducibly randomfeel free to counterargue me i neither know nor give a shit, just throwin peanuts here
>>101934127comparing quants isn't a subjective thing, the only question you should ask yourself is this one "How close the quant is to fp16?" and you can see that Q8_0 is the one with the less differences than fp16
>endless possibility at my finger tips>don't know what to prompt
>>101934141>he's literally me
>>101934141wherever i go, i must coom
>>101934141go copy some shit from /tg/, find some good prompt material there
flux knows pixel art but im not sure if it knows terms like 8-bit and 16-bit, it knows various games though and probably game systems/styles
>>101934140>8 is closer to 16 than 4hey man i solved the puzzle
>>101934123>ipndm_v+betaThat's some horror shit right there.
>>101934141Prompt nothing. Keep it empty and set randomise seeds and just explore the latents, something will stand out and give you ideas.
>>101934170you're laughing but the forge dev claimed that nf4 was better than fp8, now that's funny! https://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/981>(iii) NF4 may outperform FP8 (e4m3fn/e5m2) in numerical precision, and it does outperform e4m3fn/e5m2 in many (in fact, most) cases/benchmarks.>(iv) NF4 is technically granted to outperform FP8 (e4m3fn/e5m2) in dynamic range in 100% cases.
>>101934170it's not that simple, there's multiple "8" like fp8 and Q8_0, and Q8_0 is better than fp8 because it uses a more sophisticated quant method, same thing for Q4_0 > nf4, it's good to get those comparisons to choose the best quant relative to its size
>>101934203Who cares, both are worse than all the gguf variants
>>101934222>Who carespeople who wasted their time downloading nf4 quants, expecting it to perform better than fp8 (when the GGUF didn't exist yet), yes I'm talking about me for example :(
>>101934222holy shit are they? is that why people were sperging about gguf? i expected that to be magnitudes worse than nf4then let's see some stats niggers.
>>101934234I did too but I literally don't care it took me less than a minute to download.
>>101934165>>101934200Thanks anons, ill give it a shot.
>>101934247a lot of people are still fooled by those lies and are using those shit quants when the GGUF exist, that's not something I can respect on forge dev, he fucked it up. Now they're gonna complain that flux "isn't that good" when in reality they're using a lobotomized version of flux
>>101934243Q8_0 nearly identical to fp16. Q4_0 is better than fp8It's not that fp and nf are bad, per-se, they produce good images, but the gguf ones are just strictly better in both performance and image quality output.
>>101934250np
>>101934269>and are using those shit quants when the GGUF existyes, there's literally no reason to be using fp or nf when the ggufs exist.
>>101934168
>>101934243>then let's see some stats niggers.nf4 isn't a new thing, and it has been compared with the GGUF quants more than a year ago in the LLM (large language models) ecosystemhttps://towardsdatascience.com/quantize-llama-models-with-ggml-and-llama-cpp-3612dfbcc172(less perplexity = better)
so q8 is closer to fp16 but nf4 is better for speed at slightly less quality? or how do they differ
>>101934332look at the image again, there's everything, picture, speed, memory >>101933504
>>101934332No. q8 7zip, nf4 is like winrar, and fp16 is line winzip.
>>101934332I think nf4 is ded. q8 is better at everything
>>101934347about that comparison, just keep in mind that the anon's setup isn't very optimized, and he undervolted/limited his 3090. i have a 3070 and get higher speeds and less memory consumption with Q4
Heads up for anyone using gguf through Forge, I don't think loras are working properly, at least I couldn't get good results out of the one I tried at q8 (not compared to what I get at fp16 through comfy).
>>101934381how much vram peak do you have during inference on Q4?
>>101934391works for mewhat settings are you using? i have flux_realism_lora at 0.25
>>101934292But going by this image >>101933504 q4 is slightly slower and takes slightly more memory than nf4. Also quality difference might be subjective but I prefer the nf4 one here.
>>101934391I don't think loras are supposed to work? the weights are totally different
>>101934399>Also quality difference might be subjective but I prefer the nf4 one here.if it's not closer to fp16 then it's a worse quant >>101934140
>>101934332is it better to taste your own prompts and settings anon, I see people complaining all the time and most of the time its hardware related
pathetic
flux ain't bad. coomer version when?
>>101934396Default settings, euler beta, 1.0 lora weightThe lora was https://civitai.com/models/647940/flux-atilessence-lora-test?modelVersionId=724910Trying to produce anything like the example images results in too many wrong eyes/bad gens in general. It works but I quickly got good gens out of fp16 using their workflow so I know it's not an issue at all with the lora.
sometimes when saying pixel art, you REALLY get some old school pixel art. I wonder if "style of an Atari game" would work, will try.
Thread schizo is on his kneesRemember to report him
>>101934561>old school pixel arttell me about it
some of these samplers are kind of retarded and don't do much. has anyone generated anything good with euler ancestral or pp variants or are they just relics
>>101934577neat, even copies the CRT style
>>101934582They're not relics, they work on SD models just fine.
>>101934582euler ancestral doesn't work on flux, it was my favorite sampler on SDXL :(
>>101934586sometimes. I had something like "CRT fuzz" or "CRT distortion" in this prompt but it didn't come through reliably
>>101934582you're not making a XY plot anon? that's pretty easy to do on ComfyUihttps://github.com/talesofai/comfyui-browser
Can you use flux with Forge?
>>101934599the model knows a loooooot of gaming related stuff, even if you prompt in game screenshot of World of Warcraft, you'll get an image in that style, even with the UI if you specify it.
Ahh, fucking finally it works. Now to not make it slow as shit.
>>101934612yes, Forge supports everything (the original model, all the quants (fp, Q_, nf4) and also the loras on each quants
>>101934582Love me Euler a
>>101934612yes, just get the all in one fp8 or nf4 models (has the model + encoders + vae)https://comfyanonymous.github.io/ComfyUI_examples/flux/but for lora stuff you need comfy, dont thing it works in forge yet
>>101934623Simple as
>>101934300That's incomparable because Q4_K_M is better than Q4_0/1Idk why gguf anon chose to quantize q4_0 over q4_k_m
>>101934600checked and i am doing 10 gens with each sampler at the moment to make sure randomness doesn't have any effect. it's probably just stupid and lazy but it's thorough
>>101934620which loras work in forge? I get kmodel errors when I tried to use one
Has anyone done much LoRA training yet? This thing steals artists style better than any model I've seen yet. Like, it straight up apes their style almost seamlessly.
>>101934623me gusta
>>101934615>the model knows a loooooot of gaming related stuffmake it do half-life, tough guy
cute!
>>101934630>Idk why gguf anon chose to quantize q4_0 over q4_k_mbecause the legacy quants are easy to implement, he wanted to start with the simple stuff first, and that's fair>That's incomparable because Q4_K_M is better than Q4_0/1it's true, I should find a nf4 vs Q4_0 but I can't seem to find something like that so far
>>101934638Imagine being so miserable your only reason to exist is to antagonize a group of anons that don't want to associate with you.
>/sdg/ status:Retirement home for mentally ill avatar fags>/de3/ status:Dead and nobody showed up to the funeralHow does it feel to have won?
>>101934636Stable Diffusion 1.4 btw
>add more nodes>gen times get longer and longeranons...
>>101934625>17GB checkpoint3060 12GB bros.... we had a good run...
>>101934638What games referenced in the prompt? Flux struggles at 2.5D shooters btw
>>101934664do nf4 v2, should be fine
>>101934625LoRAs do work in forge. It even has a weird workaround for ggufs and LoRAs, but I don't really want to use forge right now.
>>101934647..was your prompt "asian woman making babies with Trump"?
>>101934168Those carrots are strikingly good pixel art-wise
>>101934674I tried a lora workflow in comfy and im not sure if its working, im gonna try it again with this random civitai lora of a character.
fucking drama shit>>101934661how? teach me. a little bone
>>101934659we are trapped in the samsara of imggen. ldg dies when theres no hipster model, sdg dies when there's no sd models. dalle dies when there's any interesting local model. the attention dies and is reborn across the ecosystem endlessly>>101934671I tried a lot to get both a half-life style and a morrowind style but flux refuses to do that era of lowpoly. it also doesn't get the direct references either. it has no idea who any half life characters are
>>101934684Schizo prompting plus LoRAs
>>101934688What if you try something that isn't a common name but instead "vortigaunt" or "headcrab"?
>>101934698god this one is kino, i want to hear it's voice
>>101934661>>101934685>>101934698These are fantastic
>>101934662yep still trying to find that sweet spot with flux
>>101934544using the blonde anime girl sample prompt>The art style of the image is reminiscent of anime pixel art, characterized by its use of bold colors, dynamic composition and pixelated texture . beautiful composition, masterpiece, best quality, beautiful girl, elf, pointy ears, long blonde hair, blunt bangs, slender, cinematic lighting, extremely detailed, small breasts, red eyes, 1girl, blush, solo, happy face, closeup, portrait, intricate dress, upper body, picturesque town illustration with moss, sunbeams, and flowers growing due to god and nature's beautyon forgedev-q8_0, euler, beta15 steps, cfg 1, distilled cfg3, no hires(ignore the black square, not sure what happened there)
>>101934699give it a shot if you wanna try but, for me, the aesthetic was so far off that I gave up. not being able to get that late 90s lowpoly style made me lose motivation
>>101934675prompt obviously STOLEN and refurbished by me (ty anon who made the original)"An old, blurry, grainy vintage, modelling photograph from a Japanese gravure magazine. The photo has a lot of signs of wear and tear. It is very degraded.The photo features a girl in a classic 70s apartment building. She is smiling, and she has classic Japanese facial features. The girl's hair is black, disheveled, and short. It is a little bit dry, and the light illuminates it softly. she is wearing vintage glasses with a thin metal frame. Her breasts are large, and they hang generously underneath the thin fabric of her t-shirt. She is wearing a t-shirt with the word "pls no" written in light yellow letters. She is only wearing white panties.Her legs are bare and crossed. Her hands are small and feminine. The girl's skin has imperfections and pores that can be appreciated in the film.There is a subtle rainbow across the lens of the camera, and the light from the sun shines softly through the curtains on a window.There are some scratches and grain on the photo. The paper texture of the photo is very evident and one of the corners is torn. she is surrounded by numerous yellow donald trump toys, in various sizes. they all look pretty silly. bleached scratched damaged torn smudged dirty wrinkled polaroid, kodak portra 400, cine film still, soft lighting"
holy kek i'm going to have so much fun with this
>>101934684https://files.catbox.moe/dc969l.png>>101934702>>101934709ty
>/ldg/ - Stable Diffusion General
>>101934725here's 0.25 strength
>>101934743>/ldg/>Stable Diffusion and any other models we want because we didn't pick a stupid name
>>101934743you guys have had sdg in your OP since your inception. why act like you're not sdg?
>>101934688we're really not, part of the main reason /ldg/ exist is to get away from you. The only reason why you're posting here now is because after weeks of spamming CP nothing has changed and your containment zone is dying. How does it feel to have our collective nuts in your mouth while you have to cope with a rentry (I made this one) where every new poster can see what level of schizo loser you are?Nobody likes you, you're the subject of pity.
Why are you using that oversized model, anon?
>>101934757kek, good one anon
>>101934742you're the best.
>>101933401So what's the verdict on Flux?
>>101934740SD could NEVER get the swastikas right for some reason, Flux is a godsend tbdesu
>>101934779it's good
>>101934779HOT SHIT BUT PATIENCE & CLEFT CHIN BRO
>>101934773these always hurt my eyes, though I appreciate your uniqueness >>101934779it listens better but is less creative. SAI still has a narrow niche they can fill: a model that follows prompts but fills in the gaps in a way flux hasn't been able to
for fucks sakeare imports broke for anyone else on Comfy's new UI?
>>101934797I pulled today and haven't had any issues
>>101934757You're laughing at him now, but SAI are the only ones genuinely concerned about making safe models, and when they crack that code they will eat the whole market. Imagine a model that cannot be tampered with to produce undesired results. Everyone will want a piece of that pie.
>>101934797>load old pic in comfy to inpaint details>no matter what seed i type in it always generates the same detailsidk does yours do the same? otherwise i think mine's broken in another wayis it "locked" to a certain seed or something because i loaded an existing pic? how do i unlock it?
>>101934664>>101934672go for Q4_0 instead of nf4-v2 >>101933504
>>101934809It's probably set to fixed in your KSampler or seed node.For me every time I drag in a new workflow it just creates a new copy of my current workflow
>>101934791>these always hurt my eyesJust zoom out mate
>>101934806I can't wait to get a truly SAFE model. the number of times I randomly see NIPPLES is OFFENSIVEand DISGUSTING
>>101934806kek'ed hard
>>101934742Thanks bro. That graininess and stippling-esque shading looks so tight, I would not know it was AI art if I saw a bunch of those images uploaded to some ArtStation
>>101934790What?
>>101934827speaking of which, how many artists do you reckon have already integrated image gen into their workflows? i bet you could save so much bullshit with a bit of inpainting, or hell you could even just sample an entire low-denoise image based on a rough sketch for you to fix up.
>>101934797i started from a fresh install and had to downgrade numpy for gguf
Is there any chance I can run Flux (dev or schnell) with my 8GB VRAM card? I have plenty of system RAM, is there any way that can compensate for the lack of VRAM?
>>101934840cutealso, im fresh installing myself soon. comfy feels like so much bloat once you have a couple of custom nodes that you use one thing from
>>101934837More than the anti ai crowd would like, that's for sure. Most will either never admitted or have no reason to tell anyone to use it. The former reason because they don't want to get "dragged" by the Twitter and Reddit mob, and the former because internet drama does not even cross their mind
if I try to use a lora my gens take 10 minutes, explain.>45.77 s/it on a 4080
>>101934850Is there any reason at all that I should use Comfi over forge (The only reason I'm using for just because it supports Flux and A1111 currently doesn't). Speaking of which, do you think Voldy plans to add support for Flux on A1111?
>>101934860Your vram is likely spilling over into ram.
>>101934860also it would be nice if I didn't have to be a turbo autist to figure out all these nodes, some use ksampler, some dont, some use lora loader, some dont, some use 2 clip nodes, some dont.
I'm seeing very heavy style variation/control with this LoRAhttps://civitai.com/models/649031/inksketch-flux?modelVersionId=726131This is not bad, even combined it with https://civitai.com/models/650330/yorha-2-gou-b-gata-2b-2b-nier-automatato get an almost dalle like result.
>>101934837Professional corporate artists, probably a ton out of requirement, but I think a lot of indie artists are too prideful/angry to want anything to do with it
>>101934869comfy has spaghetti, which is a justification for garlic bread
>>101934875>pridefulWhy is it always the most prideful and outspoken ones that produce art barely above a chris chan level?
>>101934869flux works fine with gguf models on forgedepends on whether comfyui gives you a headache or not
>>101934894talentless hacks are often jealous people yeah
>>101934894autism
>>101934894your cognitive bias. you're more willing to see the negative qualities of those whom you're predisposed to dislike
>>101934874really nice gen anon
>>101934894No true passion for art. They got into it because they wanted to be cool or get famous on Twitter or whatever. It was already difficult to do that but the AI boom made that even harder for them now that ai art floods all dedicated art sites that don't explicitly ban it (ie sites that matter). The ones that do it for the love of the game (with one notable exception I spoke to recently), generally do not give a shit about AI Edward prefer not to be part of the drama
Just ignore him I'll add a filter term to the rentry
>>101934894I don't think that's true, I've seen some pretty talented artists on X who happen to also be anti-AI and overly political
>>101934910Nta but he's kinda right. The majority of the ones being super pissy about it either do not fucking draw that all themselves or have (subjectively) low skills
>>101934837thousands, getting more every day.visual artists, stuff with projectors, ppl making the visuals for raves & parties, all kinds of things.>>101934869I uninstalled forge today, muppet. stay with comfyui. forgeman is all over the place and not someone you can rely on. voldy is dead in the water.>>101934773ahh you make those. I asked you about them few days ago. my eyes are so bad I can't get them to work lol.
>>101934870fp8 model, still slow. on a 4080. everything is fine but the second I try to use a lora, it's over 30s per iteration.
Anyone using the gguf models with ayymd on linux? I can use the regular version fine but each time I try to use the gguf ones its like it doesn't release any of the vram and chokes the system to a halt.
>>101934744>>101934725Hmm maybe something was messed up on my end? Try this one (notice the eyes here), I couldn't get it to look as clean.
>>101934941what you're describing is selection bias. its not a 'majority' of all artists but rather the demographic most willing to engage in that specific pubic opinion. you're making a mistake of abstracting a generalization from an overly vocal group that has not been proven to be representative of the population
>>101934954>I asked you about them few days ago. my eyes are so bad I can't get them to work lol.You could always use a VR headset which is the best way to enjoy stereo 3D content :)
>>101934964How it looked on Comfy, also a couple more gens (all looked horrible on forge)
>>101934957Post your task manager vram usage and the output logs when you generate with a LoRA 99% it's asking for too much vram and using ram instead.
>>101934959>ayymd on linux
>>101934971You have that proven either of us wrong though. The other guy said that the loud outspoken ones tend to be Chris chan level. We aren't saying ALL or even most people against so are like this. We're saying that the loud mouth ones tend to be.
>>101934985prompt?
>>101934980>A futuristic android wearing silver glossy attire in her station, detailed space background, looking below, classical anime style.
>>101934996it's a real anime screen kek
>>101934990read the OP before engaging with the schizo
>>101934997>A futuristic android wearing silver glossy attire is wearing goggles and bathing in a pool of blood, classical anime style.
>>10193498115.6GB out of 16, 34.20s/it, so are loras for 4090s only? this is with fp8 model and fp8 text encoder.
>>101934616>1.84s/it to 2.17s/itAt least it works I guess?
>>101934869Use comfy if you get to the point where you wanna do more things than Forge allows or get sick of having to manually go through the workflows of generating, in-painting, upscaling, etc.I started on Forge, moved onto Comfy and now I just enjoy the tinkering aspect to it.
>>101935019Forgot image
>>101935020I'm not autistic enough to make the jump yet but I'll keep it in mind
>>101935018Try loading the text encoder onto CPU. You need the room.
ran out of gens sorry>>101934978I'm a broke nigga. wait, you mean.. stereo 3d booba? zomg>>101935015tasty
>>101934869Only on comfy you can choose where to put your uner/clip/vae, whether it's on your cpu or on one of your gpu's, on and the tonemap shit aswell, that's the best thing to use if you want to do some CFGmaxxing, and I don't think Forge has implemented the GuidanceNegative, comfy is just cool if you want to get customizable features
>>101934990>You have that proven either of us wrong thoughI'm not setting out to prove you wrong. I'm saying you can't make definitive claims>tend to be>We aren't saying ALLwhen you suggest 'tend to be', that's a way to launder in an idea that "this group can be cast with this quality". the point I'm making is that you don't have the functional basis for that claim>We're saying that the loud mouth ones tend to be.I would agree with this as its the selection bias I suggested. I'm just warning against abstracting out to a larger group from this non-representative vocal sub-populationI mainly don't like when people are like "this stupid artists said something stupid on twitter, therefor all artists are like this" type shit
>>101935030how? I could gen things no problem at fp16 model with fp8 weight, and fp16 encoder, but lora breaks everything for whatever reason.
>>101935028>Warning: TAESD preview enable, but could not find models/vae_approx/Noneyou need to download this https://github.com/comfyanonymous/ComfyUI/commit/1770fc77ed91348f4060e3c0b040c1519d6f91d0#commitcomment-145464339
>>101935039>I'm saying you can't make definitive claimsI don't think either of us actually meant to. I'm just stating observations.
