Discussion of free and open source text-to-image modelsPrevious /ldg/ bread : >>101937495>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
What a horny collage
So is this just because Windows is shit?
>>1019402652 of them are mine, sorry.
>>101940291the Chinese girl and the tongue tattoo?
>>101940307Close, the Tongue and the Elf.>>101940213Here's my quick try with flux Q8 GGUF
>>101940288ask chatgpt
>>101940288Try writing it like this. That wont work on windows. d:\\ai\\flux\\
d:\\ai\\flux\\
>>101940322not bad at all, like the draft style on that one
>>101940355Fuck those are meant to be backward slashes.d:\\ai\\flux\\
>>101940288desu everytime I have that kind of errors I just copy paste that shit on chatgpt or claude 3.5 and usually they get what's wrong
>>101940322One more attempt.>>101940356Yeah not perfect but I do like the brush stroke style.
>>101940402
>>101940444Nice he looks pissed off.
>>101940444https://www.youtube.com/watch?v=6Q-97r9zbL8
How much VRAM do I need for Flux.1 Pro? I see people saying that Flux.1 Schnell is for people with less than 12gb of VRAM while Flux.1 Dev is for people with more than 12gb of VRAM but I don't see anyone saying anything about what Pro requires.
>>101940468>How much VRAM do I need for Flux.1 Pro?0 because it's not a local model>I see people saying that Flux.1 Schnell is for people with less than 12gb of VRAM while Flux.1 Dev is for people with more than 12gb of VRAMthey have both the same size
>>101940468>I see people saying that Flux.1 Schnell is for people with less than 12gb of VRAM while Flux.1 Dev is for people with more than 12gb of VRAMThey're exactly the same size. Depending on the quant you can go as low as 8gb of vram.
>>101940355>>101940368Yeah that's it. Looks like I have to clone the whole repo for it though.
>>101940369the state of /g/
>>101940444
>>101940038sorry, have a non reposted one
>>101940498We have always done things to make our lives easier. Before LLMs we were spending 5 hours to write a bash script that would save us 5 minutes.
>>101940568this
>>101940568there is a fundamental difference there that if I have to explain it to you then it is pointless to explain
>>101940578so you're full of shit I see.
>>101940578If you're not using the tools available to you to make your life easier then that's on you.
>>101940596>>101940601I see, it is pointless to explain then.
>>101940611you just admited defeat by not providing any counter-argument
>>101940611If you can't articulate or argue your position then it is a weak position.
is joycaption supposed to be slow as fuck running locally? it's barely using any of my gpu and the vram isn't over the limit
>>101940620>>101940621Like I said, if I have to explain it it is pointless to explain.
>>101940556word
>>101940636and like we said, if you don't provide any counter-argument then you lost the case
>>101940644is this a fucking court, nigger?
>You need to agree to share your contact information to access this modelAre there any torrent/magnet links for Flux dev?
>>101940650it doesn't need to be a court to know this simple fact retard, if you don't want to argue then you're just full of shit, how hard is it to understand?
>>101940650Yes.We sentence you to 2 months in /b/
>>101940671> if you don't want to argue then you're just full of shitthat doesn't follow logically but if I have to explain that to you then it is pointless to do so
>>101940667no need to go that far: https://huggingface.co/camenduru/FLUX.1-dev/tree/main
>>101940678oh it does, you claimed something, you now have the burden of proof, if you can't provide the proof then your take is dismissed and you can cry in the corner for what I carehttps://en.wikipedia.org/wiki/Burden_of_proof_(philosophy)
>>101940693my god you're retarded lmao
>>101940680Is the flux1-dev.sft just a renamed safetensors file, or is it a different format?
>>101940705
>>101940714this nigga skipped a few classes of Philosophy 101probably an American, I hear public school kids there only learn about logic when they go to college
why doesn't Comfy unload loras after applying them to the modelthe more loras the more VRAM it takes but loras aren't extra weights, they modify the weights, you can unload them after
>>101940475>>101940479Okay so if dev and schnell are the same size what is the difference between them?
>>101940800schnell takes way less steps (4 -> 8) to work but the quality is worse
>>101940800Schnell takes less steps (4-8 to Dev's 20-30) to render an image, at a significantly diminished quality to Dev.
>>101940800Dev is guidance distilled from Pro, Schnell is guidance and step distilled from Pro
>>101940812It should be said that schnel is by no means bad. Just nowhere near as good as dev
>>101940820Yeah it's a fully capable model
>>101940778so you want to overwrite the weights on the model with the lora weights? is it possible?
>>101940738>>101940714I can easily make you guys be friends again.All I have to do is go to sleep.When I wake up, the fight will have stopped, and you guys will be having fun talking to each other again.
>>101940843do I know you?
>>101940857yeah? it's me.
>>101940857Come on buddy you haven't recognized me? I'm Anon dude!
>>101940873I remember you now, you owe me $40.
>>101940892i've never met this man in my life
>>101940843Sorry but every single day there are 2 anons having a random autistic argument here, my theory is that 80% of the time it;s the same 2 anons.
>>101940938It's me, and (You).
>>101940812Nice
>>101940959A sniper that close and you will still miss me, you know why? because you love me
>>101940964very clean, catbox?
Someone was talking yesterday about a change to not get blurry preview with flux. Can't find his message anymore, someone know?
>>101940977Surehttps://files.catbox.moe/0mdz7z.png
>>101940959He deserves to die, letting a woman drive like that
I have a 12gb 3060 running on a potato system with 16gb ram, can I load and gen from a flux model?
>>101940984Don't worry the guy has a whole slew of reddit posts he shills on this general for this exact thinghttps://reddit.com/r/StableDiffusion/comments/1estj69/remove_the_blur_on_photos_with_tonemap_an/
>>101940992I have that card, and that's what I'm doing
>>101940998are you retarded? he's talking about the blurry previews on comfyUi
>>101941004yeah but I've only got 16gb ram. Anyway, it's 180secs for a 1080x1080 gen, right?
>>101940998>>101941006I think I found it again, at least the commit that implement it https://github.com/comfyanonymous/ComfyUI/commit/1770fc77ed91348f4060e3c0b040c1519d6f91d0
>>101940984it's that one anon: https://github.com/comfyanonymous/ComfyUI/commit/1770fc77ed91348f4060e3c0b040c1519d6f91d0#comments
>>101941006yeah im retarded and can't read, but i can help with that so thanks for clarifying.https://github.com/comfyanonymous/ComfyUI/commit/1770fc77ed91348f4060e3c0b040c1519d6f91d0#commitcomment-145464339
>>101941022>thanks for clarifying.you're welcome, and thanks for shilling my reddit posts for me, much appreciated
>>101941017probably notI don't know a thing about pagefile, but I assume it's extremely slower than slow
>>101941017120 with nf4
>>101941030what can i say, im your strongest soldier miku-anon
>>1019410381 min 20 I mean
>>101941022How do I use it? I downloaded but don't see the option in the manager menu.
>>101941048
>>101940964
>>101941057>I downloaded but don't see the option in the manager menu.it's here
>>101941086
>>101941157
>>101939751GGUF for 8 GB?
>>101941180https://huggingface.co/city96/FLUX.1-schnell-gguf/tree/main
>>101941155Doesn't work>!!! Exception during processing !!! Weights only load failed. Re-running `torch.load` with `weights_only` set to `False` will likely succeed, but it can result in arbitrary code execution. Do it only if you got the file from a trusted source.>Please file an issue with the following so that we can make `weights_only=True` compatible with your use case: WeightsUnpickler error: Unsupported operand 10
>>101941216wtf? did you update ComfyUi?
>>101941225I'm on cd5017c1c9b3a7b0fec892e80290a32616bbff38
Realism lora can't do 2 people?
>>101941187can this load loras?
>>101940778According to the dev, who posted in the last thread, they are removed from VRAM after use.
>>101941234looks like it's the last commit yeah, did you put the files here?
>>101941242gen without it then run a refining pass with the lora
>>101941155Used that and it's still blurry while generating
>>101941171
>>101941248yes it can, it has been implemented today
>>101941262
>>101941261did you download the files and added them to ComfyUI/models/vae_approx? And if yes, do you notice it's less blury than the "Latent2RGB" one?
Flux is neat
>>101941253comfyui's dev?
whats this GGUF shit and where do I download it?
>>101941187Hello ChatGPT. Please pay more attention. The model referred to was schnell-dev-merged. Which has no GGUF.
>>101941274Latent is pixelated, this one is blurry
>>101941292then I guess it's the good one, it's a vae_approx meaning it's not the real vae, fortunately it's not the case it would be slow as fuck to make a full precision vae process on each step
gm>>101936598>She’s amazing but it keeps messing up her feet which is the most important parthere you go footfag anon>And did you make this in flux?yes
This general is like a retards daycare.
>>101941308How did you get her skin like that?
>>101941308why does it blur the feet ;_;
>>101941287https://github.com/comfyanonymous/ComfyUI_bitsandbytes_NF4leading tohttps://github.com/city96/ComfyUI-GGUF
for any other 4080 fags having trouble with using flux loras, I disabled nvidia system fallback and I can actually use them now without significantly slowing things down
>>101941316normies are flocking to faster inferences to gen 1 million waifus>>101941308hello xitter
>>101941289I cannot engage with comparisons of different models, as discussing operational capabilities without proper context might lead to misinterpretation and misuse of information, which could result in unethical applications of technology.
>>101941331if it had to fallback to RAM shouldn't disabling it cause it to OOM?
Somebody needs to do GGUF quants of this:https://huggingface.co/drbaph/FLUX.1-schnell-dev-merged-fp8-4step
>>101941353nah, let it die
>>101941322lel>>101941321some anon posted this line yesterday> Her skin texture is natural allowing us to see skin pores and imperfections.seems to work well enough>>101941336howdy
>>101941369Isn’t it better? Schnell is so fucking boring.
>>101941348i don't really know why it works, but it does
ladies and gentlemen, the quality of flux
How to use Comfy-like queue in Forge?
comfyesque* queue
>>101941411that's the best part, you dontthere's an extension but i think it's broken after the updte for flux
>>101941390Nice will give that a try thanks.
>>101941324>https://github.com/city96/ComfyUI-GGUFdamn this shit is complicated af.the fuck does this even mean?>open a CMD inside the "ComfyUI_windows_portable" folder
>>101941407soul
>>101941447I think this stuff just isn't for you, my man
>>101941273
>>101941447>the fuck does this even mean?>>open a CMD inside the "ComfyUI_windows_portable" folderwhy won't you ask bing chat for that kind of stuff? that's what I do and I learned a lot with it
>>101941468
>>101941447>damn this shit is complicated aflike fr fr my guy, me and skyler just tryna gen some skidibi. anyone know if i can do this on my ipad?
freaky
>>101941431Why it's the best part when users, for example, cannot run several XYZ tasks at once? Forced to wait for the previous task to finish.
>>101941496would
Maybe I'm fucking up but as a 8 vramlet gguf seems way slower than nf4
>Try Flux>It's most of what I could ask for for a base model>Train LoRAs>They're shockingly good for the effort put in>Controlnets are even available>Still find myself getting bored.Guess it's time for me to go into hibernation and wait for finetunes and optimizations.
>>101941473>>101941462>>101941488so which one do I need?
>>101941517Q8? Yeah it takes up more Vram, I think it's best for 16gb vram, but 12gb is ok but a bti slower.
>>101941535depends on your machine, how vram does your gpu have?
>>10194155524GB
>>101941517Also love her legs.
>>101941559then go for Q8_0, it's the one with the closest quality level to fp16
>>101941559don't go for Q8, just run FP16
>>101941548When video games used to be cool
>>101941544Q4_0 actually
With 12GB VRAM and 32GB RAM, which quant or fo should I go for for max speed if I want to use LORAs and for now don't want to mess with CFG >1 (does this affect RAM usage even?)
>>101941574but muh LoRas, all that shit aint fitting into my 4090
Challenge for you guys: A person slicing something with lightsaber. Like a car or a person.
>>101941586loras don't take extra space, they modify the weights. if the interface you're using takes more VRAM because of loras then it is shit
>>101941586this, fp16 is too big if you wanna add some stuff on top of it (loras, controlnet...)
>>101941597How dare you slander comfy like that.
With lora flux is starting to get good at nsfw prompt. Genned a few changing room footage with natural pubic hair.
>>101941596think that might be a spiritual successor to this
can you use FLux with Automatic1111 yet?
>>101941621and you don't catbox one for us? you're a terribly selfish anon
>>101941629A1111 is dead anon, Forge replaced it
>>101941641Did the Ukrainians get him?
>>101941638My ip pool is small. I used 649568 and 651940 from civitai.
>>101941641Stop spreading FUD. Voldy is the best.Comfy and Illyasviel are angry bickering troons.
>>101940475If Pro is not a local model, how is there a download for it at Civitai?https://civitai.com/models/618692?modelVersionId=699332
>>101941686
>>101940556neat 1girl
>>101941710>Voldy is the best.He's so good he still hasn't implemented Flux 2 weeks after its release
>>101941721it's 'training data', not an actual model
>>101941721Idk, and I think it's retarded, it's not a local model it shouldn't be there in the first place
>>101941517Fellow 8 vramlet here. Do you use ComfyUI_bitsandbytes_NF4 with bitsandbytes and this model? https://huggingface.co/sayakpaul/flux.1-dev-nf4I want to try anything that might make this go faster.
>>101941730heckin BASED, we are so back a1111 sisters
flux compatible lora isn't working with gguf dev0q4, it's having no effect at all. I've updated gguf today on comfy.Something obvious ive missed?, workflow is at defaults, only model and loader and lora have changed been added
bitsandbytes is not usable on ROCm is it?
Is using this with LoRAs a viable option for a VRAMlet?https://huggingface.co/OlegSkutte/SDXL-Lightning-GGUF
>>101942037It is usable, almost always was, but always needed a fork, lot of them went unmaintained but AMD took it upon themselves to maintain one https://github.com/ROCm/bitsandbytes
I'm making progress, finally got the LoRa shit to work.
>>101942110Thank you. I'm going to try and see if nf4 runs a bit faster than gguf on 8 GB.
I just realized i want to make wallpapers with some nice gen settings i have, What settings would i be changing to achieve 1920x1080 and other potential wallpaper-y resolutions?>thinking something for my PSP as well for the fuck of it would be rad so i may be doing weird resolutions
>>101942037you don't need to torture yourself with nf4, go for q4_0, that shit is the same size and better quaity >>101941569
>>101942171Also you can try their upstreaming effort https://github.com/bitsandbytes-foundation/bitsandbytes/tree/multi-backend-refactor, would be nice that after more than a year to have it a HIP backend merged.
>>101942183If it's faster, I don't care about quality, really. This thing is painfully slow.>>101942196thanks.>Could not detect model typesigh
Is there a guide on settings for webUI?I probably don't need to touch them much, but there are so many of them…
good quality/quantity ratio for lora training?i have literally thousands of decent period-appropriate pics, many (but not all) of which contain the thing(s) i want to train onwas thinking of running them all through joycaption and starting training, but would that pool then be too broad to capture anything of value? should i reduce it to 10-100, 100-250 pics or something else?on a side note, how important is captioning anyway if i just want to enhance the detail on something the model can already generate?
>>101942217more is always better especially if it captures a diverse variety of descriptions, the most important part is ensuring you have captions in each that capture the purpose of the lora, ie if you're doing vintage pictures, you should make sure, vintage, (year), etc are in each caption.
>>101941064get new material faggot
>>101942183>>101941569>better qualityBased on what? Your image shows less RAM utilization, more speed, and variations in the image compared to the others that would be no different to using a different seed. Prompt adherence is basically the same as the other lower quants (it ignored the color of the skin, that's it.)Your herd mentality is tricking you into believing standing out is a problem.
>>101942301>Prompt adherence is basically the same as the other lower quants (it ignored the color of the skin, that's it.)lol, nf4 is the only one that has Miku making a different pose than the others>Your herd mentality is tricking you into believing standing out is a problem.You have no idea what the goal of a quantization is do you?
I don't want to learn new diffusion stuff
I'm tagging dataset that's made out of ultra painterly paintings and all taggers and vllms make at least 3+ mistakes per image. Shit takes like 2-3 minutes per image to fix it
I have a 6gb vram and 16gb ram. Can I even run Flux?
https://github.com/intel/AI-PlaygroundIntel beat nvidia and amd to the punch
>>101942371if it takes 2-3 minutes just write the captions yourself
>>101942318Based! tiny white peckers btfo!
>>101942325>the only one that has Miku making a different poseAnd this makes the NPC uncomfortable.
>>101942318I used to gen a ton of these, kek
>>101942373Why bother? It can't do porn. Not well, at least.
>>101942447it got decent nude loras faster than XL
>>101942447It's cumming....https://civitai.com/models/645697/facesitting-flux?modelVersionId=722322https://civitai.com/models/655732/cum-on-face-flux?modelVersionId=733630
>>101942373On comfy, yesI have yet to find a workflow that gives good results >>101942458It didn't, and are locked behind a patreon, fuck You.
>>101942480>It didn't, and are locked behind a patreon, fuck You.it did and I'm talking about the available civitai loras, which patreon loras are you talking about?
Another flux lora creator training on AI images (pony)WTF is wrong with these people.https://civitai.com/models/653944/female-solo-masturbation-with-pussy-lips-spread-fingering-while-on-her-back-with-legs-up-pov?modelVersionId=731604
>>101942373works on my 1660s with forgeflux-schnell-nf4, 4 steps768x768 took 1 minute1280x720 took 2 minutedon't try flux-dev, 20 steps is too much
>>101942447There is some alright lora like:https://civitai.com/models/649568/asian-female-pubic-hair-and-armpit-hair-for-flux1-devhttps://civitai.com/models/651940/hidden-spycam-changing-room-fluxhttps://civitai.com/models/649762/onoff-for-flux-onoff
>>101942196What about this one? It's supposed to be ported to HIP:https://github.com/agrocylo/bitsandbytes-rocm/Unfortunately, none of them work. It's not like there was going to be much improvement, seeing as I get 40 seconds per iteration, but I wash willing to try.
>>101942513Please understand, there simply no footage of masturbating women on the internet.
>>101942383It varies, sometimes I just have to edit it for 20 sec, sometimes I will be there for 5 minutes because even I have problems describing stuff in the image lol
>>101942526you can do on/off just with prompting but the lora does bring some understanding of nude bodies
>>101942537Oh my bad, understandable. Carry on.
>>101942513>RubberDollfetishman69 has entered the chat
>>101942529As I said, there is multiple fork that got unmaintained with time. This one last commit is more than a year old and was because text model used bitsandbytes for inference for a while, once they stopped using bitsandbytes for text gen, that fork stopped being updated.
Wait, do LoRA's work with schnell too, or only dev?
>>101942342Too late anon.
>>101942578>You're Sinoneway to dox that anon man.
>>101942565They work with bothhttps://huggingface.co/XLabs-AI/flux-lora-collection
>>101942606Try this one too to comparehttps://civitai.com/models/650574/andrea-botez-flux
>>101942632
>>101942663
>write side-by-side clothed and nude images of a young girl with the side by side lora>an image starts to form in the preview box>awshit.jpg>kill process, close browser, clean caches and temporary file storageHoly shit flux
>>101942663Same prompt as this but this image turned out...disturbing. So Catbox.https://files.catbox.moe/xatun6.png
Good node for stacking an arbitrary number of LoRA?
>>101942701rgthree's Power Lora Loader
>>101942715This is better than the preview images
>>101942712That one has only up to four. I was thinking one that could have an unlimited number of rows, if that is possible with the UI.>>101942681Journos are going to have a field day with Flux once people catch on how easy it is to generate certain kinds of illegal porno with LoRAs that have nothing to do with it.
Who the fuck uses thishttps://civitai.com/models/647946/pyros-slurp-long-tongues-for-flux
>>101942751>That one has only up to fournot the Lora Loader Stack, use the Power Lora Loader, you can add as many as you want
>>101942760Imagine blowjobs…
>>101942741was a lucky seed, some of my other gens with the same parameters haven't looked as nice
>>101942715catbox......?....very....................................nice.........oh my......
