Discussion of free and open source text-to-image modelsPrevious /ldg/ bread : >>101957363>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>>>/g/sdg>>>/h/hdg>>>/e/edg>>>/d/ddg>>>/b/degen>>>/vt/vtai>>>/aco/sdg>>>/trash/sdg
>>101959699>I made it into the collageMy life is now complete
>>101959721why am I not in the collage tho
>>101959761what happened to your fingers bro
>>101959761Why are you samefagging in the other thread schizo?
>>Maintain thread quality >https://rentry.org/debo
The new rf variant for dpmpp_2s_ancestral sampler in comfy is pretty good. Of course it's also half the speed of euler, which kinda sucks.
>>101959788>Posts malware in his news posts for weeks and tries to shift the blame onto others for calling him out for it, under a changed filename but forgets to change his stylegod damn
>there's a webcam nodeDoes this work with animatediff?
>>101959725funnily, adding tonemap made the speed inference faster, went from 7s/it to 5s/it, more speed is always welcomed
>>101959799If each step takes twice the time, then it needs half the steps overall
>>101959823ok nvm it went back to 7s/it :(
>>101959823There was a reddit link with a bunch of guides, including using tonemap. Is it in the OP somewhere?
how do I change the guidance number in forge
>>101959721I made it 2 times in a row, yeah baby!
this shit https://civitai.com/models/657422/flux-paizuri ALMOST works https://files.catbox.moe/j0ml4n.png
Don't skip leg day, fellas.https://civitai.com/images/23837590
>>101959836this? https://new.reddit.com/user/Total-Resort-3120/submitted/
>>101959836https://old.reddit.com/user/Total-Resort-3120/submitted/One day this'll be in a rentry and I wont have to link to a reddit account
>>101959825Maybe the lesson learned from this is "just double your steps for better results".
>>101959799try the sampler called "supreme" I think it might be the besthttps://github.com/Clybius/ComfyUI-Extra-Samplers
>>101959799>Of course it's also half the speed of euler, which kinda sucks.that's why I use Deis, it's better than euler but the same speed
>>101959850SD3 lora, porn included
>>101959855>>101959856Thanks.
>>101959862Is it slower?
>>101959862>noted to my google chrome bookmark, so much shit to learn goddam
>>101959888I don't think so but I'll have to double check.
>>101959862>>101959875>>101959895So much to learn. I exclusively use Euler A lately since i've been fixated on Pony, would love a faster sampler that achieves the same whatever the fuck Pony needs from Euler..
>>101959888okay my bad it is 50% slower than euler.(euler 14 seconds, supreme 21 seconds)
The homicide detectives looked inside her phone, among the thousands of selfie images she had taken the last one she took stood out to them. A vital clue to the case!
>>101959953>I exclusively use Euler A lately since i've been fixated on Pony,Euler A doesn't work on Flux... it was my favorite sampler
>>101959973What did you settle on? Also schnell or dev?
>>101959725>Also tonemap seems to be compatible and adds more detail on top of the PAG gen. This is attonemap 0.42.It's also compatible with Dynamic Thresholding, noice!
>>101959973Good, flux is forcing forward progress for better faster stronger samplers. That makes me even more optimistic for when 2D is further improved.
https://x.com/realDonaldTrump/status/1825138139502878806fluxgods...
>>101959979dev
Why has Comfy have no option to pause or save the queue so you can do something else and resume later?
>this is what that chicken gen looks like in diffusion toolkitwhaaaaaaaaaaaattt thee fuu-我爱北京天安门,天安门上太阳升;伟大领袖毛主席,指引我们向前进。
>>101960002at least, if orange man will be elected, he won't for sure make AI deepfake illegal kek
>>101960006But what sampler did you settle on?>>101960002American politics is so weird for someone from Europe, where an actually nuanced spectrum exists.
>>101960002omg someone could think this was real
Another blonde victim which matched the previous victim's description and job (influencer) also had a similar selfie photo found on her phone, could this be the beginnings of a serial killer case?
>>101960021>But what sampler did you settle on?oh, Euler, and from time to time Deis, the others are or else shit or else too slow
>>101960021To help explain it better, american politics is like choosing your football team, once you choose sides the nuances don't matter, you win or lose. Everyone else is the ENEMY.
>>101960021yes in europe you get to pick from 50 shades of socialism instead...
>>101960019He did have a moment where he was targeting violent video games...so you never know.
>>101960117Not the hecking moment?!?! He is fascism y'alls
>>101960110This was my point. Everything in the US looks like different flavors of neoliberalism (with or without religion) from here. Off topic tho.
>>101960008I thought of a pretty good idea recently where you'd have two queues: a queue for exploration and a 2nd queue for upscales.So whenever you find something good, you could queue an upscale of it, but this upscale queue would only run during downtime (like when you're writing prompts or alt-tabbed).It would basically always be running automatically during any downtime. And so it should also pause/interrupt whenever you run your main exploration queue.
>>101960031>>101959966The third incident ended with the death of the killer by the hands of his victim, this case ended without becoming a serial killing case which the higher up were very happy about.>>101960177Very nice
>>101960221nasty coom cave
>>101960243So he wore dreadlocks to try and decieve the detectives into thinking he was black.
>>101959875what settings do you use for DEIS? It's waaaaaaaaaaaay different from Euler. All my gens are weird and kinda deepfried with it, regardless of cfg scale.
n slur
>>101960297>what settings do you use for DEIS?what do you mean? I just chose Deis and I don't have other options
>>101960002Based.
>>101960322>I don't have other options
>>101960335deis normal 30+ steps I guess, beta + deis gave me weird shit yeahhttps://reddit.com/r/StableDiffusion/comments/1em7zy3/testing_the_samplers_schedulers_with_a_xy_plot/
Ok, seems like PAG can be tammed, you have to crank up the scale to fucking 30 at least kek
>>101960348I think a lot of my samplers might just be broken, given i tried those identical settings to start with and it wasn't great.DPMPP3MSDE has been broken since i started using comfyUI, always worked in forge.
>>101960373anon, I hope you're adding Dynamic Threshold or Tonemap in addition to CFG = 6 right?
>>101960277You'd think so but in fact it was a "skinny whi boi" white body-suit!
>>101960393never been told i needed either of those things when i asked a few days ago show me a node setup so i can do it right pls
>>101960407Here's a tutorial for Dynamic Thresholding (good for anime, not so much for realistic shit)https://reddit.com/r/StableDiffusion/comments/1ekgiw6/heres_a_hack_to_make_flux_better_at_prompt/And here's a tutorial for Tonemap (Great for realistic shit but not as good as Dynamic Thresholding for anime)https://reddit.com/r/StableDiffusion/comments/1estj69/remove_the_blur_on_photos_with_tonemap_an/
https://github.com/comfyanonymous/ComfyUI/discussions/4457You guys are downloading the latest standalone release, aren't you anons? You only have yourself to blame if performance is bad.
>>101960429well in this case its a funny amalgamation of Pony re-utilizing a realistic style, usually Euler A is fine but my old better gens on forge were with DPMPP3MSDE, with a higher cfg scale. Normally i work with 3.0/3.5 and Euler A.Ill have to check that out for the normal workflow.Now im back to fixated on why 3MSDE isnt working.
>>101960402That's what they thought, but the autopsy revealed a Jamaican man underneath the white body suit.
>>101960443he says download the latest standalone AND THEN UDPATE.This is just so people are using a clean repo when reporting issues.
if you have enough VRAM to fit the flux fp8 checkpoint version, would nf4 or q4 or whatever still offer better throughput? got 2.24s/it on the fp8 checkpoint version. i also have a q8 quant which at first glance has a lot worse quality than the fp8 checkpoint and is slower too
https://civitai.com/models/659438/xenomorph-egg-flux
>>101960262ya Beksinski lora makes good slime
>>101960491Kek
>>101960491LMAO
>>101960483>if you have enough VRAM to fit the flux fp8 checkpoint version, would nf4 or q4 or whatever still offer better throughput?why not go for Q5 or Q6 instead?
Have you found a consistent way to prompt for young girls? If I go with "young girl", they're often WAY too young. And I if say something like "20 year old girl", or worse yet, "woman", they look in their 30-40s.
>>101960491lmao
>>101960528Try Young woman
>>101960511ok sure, but can i expect better compute throughput from a reduced model or only if it helps meet VRAM compatibility?
>>101960491ok, now that's funny.
>>101960539Huh. I'm so used to SD1 understanding age ranges so well that I had not thought about this combo.
>>101960574Flux sure does look good, but it's spatial understanding is garbage.
>>101960594I know, it's like the last one I tried too.
First time lora training attempt, what does "steps" actually mean? If I'm training 5000 steps @ rank 64 on, say, 20k images with a batch size of 1, does that mean it won't even look at 15k of those? I guess raising batch size could help but I've heard raising the batch size lowers quality...If it matters, I'm not trying to capture a general style, but rather specific items and their associated details. Currently Flux just generates a melted/distorted mess when trying to generate those details.
>>101960741>>101960753what's this guy's problem?
>>101960528yeah its tricky. flux extremely sensitive to prompt. hyper-sensitive. might end up with a toddler. "cute naked (cough) girl on a bed", I got a baby. >>101960719sick
Okay, I'm doing some tests to see what Comfy parameters give me the best performance, and I set up this simple workflow. But something weird seems to have happened here. What am I missing? Why is there a dude here instead of the bottle?>>101960767Some things about this model are very German.
thanks for OOT lora anaon
>>101960612Flux is BFL's SD 1.5 moment, and it only gets better from here for future iterations of this stuff. And it's actually being done by a competent team for once.
>>101960794I used the Force/Set Device to move the VAE and CLIP to the CPU and it's producing a bottle instead. How bizzarre.
>>101960766Working at Mcdonalds for so long does things to a peron.
https://youtu.be/PrPY_USgngg?si=2upQLT9Es6loGyLd&t=202uh-oh, niko is being a disgusting ai-bro again and stealing artists jobs
>>101960818What does "BFL" mean?
Can i run flux on 1080ti?
>>101960835Big Fucking Loseror maybe Black Forest Labs instead
>>101960835Brown Fluid Lavatory
>>101960845Ah, right. I'm only vaguely aware of things. They used to work on the original SD, right?>only gets better from hereThat's exciting. I might end up buying a new PSU and a 3090 towards year's end.
>>101960794you are missing the model sampling flux node or am I blind? don't ignore it. thats why it doesn't pick up your idea. also the guidance is an important value, too high will of course fry the image but it needs to be constantly adjusted.what are you tring to imply about us here? son.
>>101960873>model sampling flux nodeI just use the default KSampler. It works (pic related). But I don't understand why it rendered a man instead of the bottle in that first picture. Other than moving CLIP and VAE to CPU, it's exactly the same setup.
>>101960901just squeeze the "model sampling flux" node between unet loader and ksampler and try it out. and of course you need the image width and height for that.
>>101960860>Ah, right. I'm only vaguely aware of things. They used to work on the original SD, right?yes, a lot of the BFL crew are former SAI.As for a 3090, that's a pretty good purchase for 24GB vram at a decent speed for image generation. Just make sure you pc case has good cooling as these cards get toasty.
https://enragedantelope.github.io/Styles-FluxDev/>600 styles test>they all look the same:(
>>101960999https://civitai.com/images/23199531>Style Test Workflow>Not using any of the negative guidance hacksNow there's your problem
>>101960971What's that for?
Mexican weather lady using OOT lora
>>101960971What is ModelSamplingFlux for? It adds 2 seconds to my generation time and it generates the exact same image if I don't use it.
>>101960760A step is essentially the batch size number of images processed. So 5000 steps with 15,000 images does mean 10,000 images aren't looked at. The performance gains of batch size far outweighs any quality loss (made up with more steps or slightly higher learning rate). With Flux you can't do more than batch size 1 on consumer GPUs. Generally you can expect to do 2-3 epochs to get the results you want.
>>101961060would play
>>101961032>>101961094copy pasted straight from reddit but very much relevant:"base shift is a small, consistent adjustment that stabilizes the image generation process, while max shift is the maximum allowable change to the latent vectors, preventing extreme deviations in the output. Together, they balance stability and flexibility in the image generation.Using a dog as an example:Increasing Base Shift: If you increase the base shift, the generated image may become more consistent and closer to the intended form (a clear image of a dog) with less variation or noise. The dog might appear more stable, with well-defined features, but it could also lose some subtle details or become slightly repetitive in texture.Decreasing Base Shift: Reducing the base shift could introduce more variability, allowing for finer details or more nuanced textures to emerge. However, it might also make the image slightly less stable, potentially introducing minor artifacts or inconsistencies.Increasing Max Shift: By increasing the max shift, the model has more freedom to explore the latent space, potentially leading to more creative or exaggerated interpretations of the dog. The dog could end up with more exaggerated features or a more stylized appearance, but it might also risk deviating too much from a realistic representation.Decreasing Max Shift: Lowering the max shift would constrain the model, leading to a more controlled and realistic depiction of the dog. The image would likely remain close to a typical dog appearance with fewer unexpected variations, but it might lack some creative elements or subtle uniqueness."tldr "artistic freedom slider"
>>101960999>>101961030>Abstract Painting Winston Churchill looking at iPhone., Painting in abstract styles, often using non-representational shapes, colors, and forms.People really need to stop sleeping on CFG 6~ and high Guidance Neg.This entire test needs to be redone with new parameters lol
downloading these quants at a mere 8MB/s just to compare s/it and schizo interpret render quality is annoying, especially when they all just seem worse in every way than the regular fp8 checkpoint
>>101960794Don't bother with gguf
>>101961183I can't use anything else.
gu gu gu gu
>>101961174this dude complaining about having 80Mbs, I have 20Mbs at best and it drops at night.
>>101961200no, huggingface is only giving me that much
>>101961172>CFG 6Gen time is ~3 times slower on my poorfag gpu.
>>101961210then you're an even worse of a complainer than i thought
>>101960999>>101961172Compare that to the output from the test (pic rel), which doesn't use any guidance neg - and you can clearly see that there's so much potential in Flux that is wasted by not using negative guidance.
>>101961219yeah, for artistic style, Dynamic Thresholding + CFG 6 is really good, I wish DT would be as good on realistic images though, Tonemap has the opposite problem, imagine if there exist a method that would take the best of both worlds, would be fucking perfect
I've been out of the loop for a while, its forge abandoned now? Do I need to switch back to automatic1111?
>>101961239woah
>>101961214Yeah it's double for everyone at the very least since it generates another noise latent to compare against at every step, so the negative attributes can be avoided. (At least that's how I believe it works).In related news I just updated to Pytorch 124 nightly and my gen times on Flux reduced by 10 seconds lmao.
>>101961239The original Matrix is so post-processed to shit that Flux images look as if they were from the film.
>>101961183q8 gguf is almost indistinguishable from fp16, unlike fp8...i've gotten better results and better text when I started using q8
>>101961217not my fault you're trapped in some godforsaken hellhole like australia
>>101961274Same here
>>101961266If you guys wanna try that yourselves the command I used on my python_embedded folder is.\python.exe .\Scripts\pip.exe install --pre --upgrade torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu124
>>101961274>>101961286Not the schnell one
>>101961270Lol true
>>101961296I'm the second anon. I'm using schnell.
>>101961274>q8 gguf is almost indistinguishable from fp16thats bs and please dont post the miku with dreadlocks image as an answer
>>101961219with all due respect.. the first image looks like dogshit. its not abstract at all. and the 2nd one is just fried garbage. maybe try a different approach?>>101961239fuck
>>101961319nta but this is widely known and the reason quants are a thing at all. 16 -> 8 is indistinguishable
>>101961317Try this and compare it with gguf schnell q8https://huggingface.co/drbaph/FLUX.1-schnell-dev-merged-fp8-4step
g guff
>>101961249please respond
Pretty crazy how well flux does with low resolution, this kind of low resolution before would have totally messed up faces that I would have had to inpaint over.
>>101961324>the first image looks like dogshit. its not abstract at all. and the 2nd one is just fried garbage. maybe try a different approach?you can't do anything about it; flux just doesn't know that concept
>>101961337>schnell-dev-merged
>>101961363No Forge dev is who was actually what is giving everything a boost, making everything work easier for everyone with different GPUs and it all works with Forge.so it's better than ever right now, A1111 is pretty much dead (for now)
guys help plsIm getting all slop results on forge, I try and follow all prompts from popular images on civit but my results dont match at all!My gens are always yellowish, noisy, grainy, foggy and got some artifacts, not sharp at all, the results here and on civit look sooo fuckin good! Im on forge, using the NF4 version of Flux3070 and 32G RAM
>>101961332>this is widely knownthat may apply to text LLM models, we're just discovering image quants models, I tried the Q_8 model after some posts stating that it was closer to fp16 than fp8 and all it gave me was some shitty results compared to f8quants model seems like some vramlets cope so they can boast about smaller models without able to test properly against bigger models because they don't have the proper hardware to start with
>>101961324With all due respect, if you actually read the posts then you'd realize that the second 'fried garbage' image isn't mine - it's from this "Styles Test" another anon posted.https://enragedantelope.github.io/Styles-FluxDev/I was using the same prompt, and quickly popped it into my workflow.My gen was used as an example that base Flux has a lot more potential in it's ability to output unique styles that most people aren't aware of.Cheers.
>>101961391show screenshot of forge page
>>101961391nf4 and gguf are slop
>>101961374try "abstract art", "painting", I dunno.. "modern art", whatever and an abstract artist in the CLIP_L box
>>101961363Have you tried using your eyes.https://github.com/lllyasviel/stable-diffusion-webui-forge
>>101961416here>>101961423ahhh damnn
>>101960794Looks like it's fucked with safetensors too on latest comfy, the text encoder just outputs garbage.
>>101961454Have you tried Euler with Beta and without highres fix?
>>101961480bitch
>>101961485Ive tried euler with and without hires, but not beta, thats something I gotta run onceIll post a gen is there no hope really for 8G plebs?
>>101961426but what about https://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/801 ?it says to go back to the main branch which links to automatic1111
>>101961516I think you need to download ae.safetensors and put it in the vae folder then load it at the top.https://huggingface.co/black-forest-labs/FLUX.1-dev/tree/main
Reminder
>>101961319Are you using dev or schnell? All I know is that q8 dev is virtually lossless compared to fp16 dev. This is probably different with schnell though, being much less steps.
>>101961530ohhh I didnt know this, I thought flux had no VAEForge allows it?>>101961485picrel is euler beta 35 steps no hires, 1min40s
>>101961556take the comfypill
>>101960443This dumb motherfucker>Even 4090 doesn't have enough vram for that YEAH NO SHIT SHERLOCK FUCKING IMPLEMENT BETTER MULTIPLE GPU SUPPORT YOU FUCKING WANKER
I got 2 GPUs, can I choose to load flux on the 3090 and load the VAE on the 3060 using comfy? Everytime I generate an image it loads flux (takes about a minute or two) and then has to unload it to load the VAE in the same GPU, so it takes at least 3 minutes to generate an image
>>101961556>35 steps10's enough for most cases unless you like the gen
>>101961538is this that flux thing all the kids have been talking about in school?
>>101961520Even the latest update in that announcement is before flux release. He has been committing tons of code each day for a while now. I don't know if he will pause again for whatever reason.There is a forge fork that might be active if he does.https://github.com/Panchovix/stable-diffusion-webui-reForge
>>101961124I could run training with batch size 4 on 12GB (at 512x512)
>>101961596Things are in motion everything is in flux.
i have a folder full of pics with genned captions and want to create a lora from themsince training seems to be in such a state of flux (heh), what's the DEFINITIVE latest&greatest way, as of today, to train with best results?
>>101961582I did, but for some odd reason its slower than forge, but its much faster for XL PDXL and 1.5>>101961591hmm understood.what about cfg and distilled cfg? should it always be at 1 and 3.5?
>>101961610frfr
>>101961556This looks fine
>>101961616I used civitai's trainer, it's like 2 dollars to train a lora on flux and it turned out pretty good for the first time in my case. I could use my own GPU but Im too lazy tbqh
>>101961617>>101961591>>101961582also anons is there some magic prompt format for Flux? Im a SAI promptschool graduate, know fuckall about Flux>>101961633damn anon this is so much better! Is this from the main model? the non nerfed nf4?
>>101961640>damn anon this is so much better! Is this from the main model? the non nerfed nf4?looks like pap you blind mongoloid smdh
What's a GPU weight slider in updated Forge? Couldn't find anything about it.
>>101961636meh, i've got a 3090 and the patience to run it locally, just don't want to pick the wrong tool to start with only to find out 10h later that something else was better
>>101961640>also anons is there some magic prompt format for Flux?Just use this, if you're not into writing essays https://huggingface.co/spaces/gokaygokay/FLUX-Prompt-Generator
>>101961585there was a custom node for that, mikufag linked it before
>>101961656rude
>>101960348Here's a comparison of the samplers using the simple scheduler. heuncpp2 converges the fastest but is also the slowest. But other samplers start pulling ahead at the 10 step mark. I find also and tested that ipndm gives the best from 15-20 steps if you look at the pie and background details before deis takes over at 20-30 steps looking at the pie again and then every scheduler converges above that step count. Mostly using ipndm at 15 steps for that reason for my gens and deis if I want something better. I like ddim_uniform better so I am using that as a scheduler instead of sgm or simple.
tell me, why should I use Forge and why will I get better results with it?
>>101961667In that case SimpleTuner will do the job, are 24GB of VRAM enough to train though?
>>101961668thanks, if someone can respond about cfg and distilled cfg, and comfy being slower than forge, it'd be great..also am I missing anything else I should know about?
>>101961683gross, furries should navigate to /trash/
>>101961667you won't get it right on the first trybe prepared to spend a lot of time on it, refining and improving
>>101961686Tell me what did you eat today and why?
>>101961669where?
>>101961703I had some hotdogs and I just felt like itit's a nice sunday so a bit of unhealthy food is fine
For AMD users on older cards that don't support WMMA, I made a ComfyUI node that implement flash attention in triton.It's sadly slower than sub quadratic or split cross attention, but it uses way less VRAM, might be useful for some.https://github.com/ardfork/ComfyUI-flash-attention-triton
>>101961696Maybe you need the T5 text thing too look at this thread it should helphttps://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/1050
>>101961505what the duck?
>>101961717Sounds real good.
why should I use Forge and why will I get better results with it?
>>101961633>>101961556>>101961530>>101961668ok so this is with ae.sfteuler beta and prompt from the generator.is there a way to run that prompt generator locally?
>>101961696>have to link to reddit now>https://old.reddit.com/r/StableDiffusion/comments/1en5eb3/flux_realism_lora_comparison_onoff_and_different/You can ignore CFG for now. >Distilled cfgLower (1.6-1.8) for realistic images and higher 2.75-3.5 for normal ones. Higher guidance (dcfg) also means less errors/artifacts
>>101961690think so, i've seen others have success with 24GBsupposedly it's the bare minimum though so not optimal>>101961701eh, i'm mainly just concerned about getting the method right first try, since flux is relatively new and it feels like there's new discoveries, tools and tricks being posted every daydataset quality/config is another matter and i don't mind having some trial and error with that, hopefully if i curate it a bit beforehand it shouldn't be too bad though
>>101961798thanks anon, Ill go thru this>>101961734I thought the t5 thingy was for the main models only and not for the nf4 ones
>>101961716dude needs a rentryhttps://old.reddit.com/r/StableDiffusion/comments/1el79h3/flux_can_be_run_on_a_multigpu_configuration/
>>101961789>is there a way to run that prompt generator locally?https://github.com/Ling-APE/ComfyUI-All-in-One-FluxDev-Workflow get this workflow if you want everything in one place and read the page for instructions
>>101961912their husband will be furious
>>101961912delete this
>>101961921Funnier than what i was gonna say lmao
>>101961835afaik, nf4 has that and clip_l baked into the file but you can also supply them separately
>>101961959that would increase vram usage Im guessing, I think Ill get an oom error
>perhaps it's not perfectwell only one way to find out :^)https://civitai.com/models/659046/mistress-of-the-flux
>>101961977no idea, i just select t5 and clip separately and have no issue running it on my 3070, i think it just loads in the weights for the main model only in that case
>>101959823>funnilyRemove this word from your lexicon please, my inner dialog finds it's sound offensive. Humorously, interesting or counter intuitively would have have been better options.
>>101961912
>>101962020fffunnaallaayyy
>>101962015would you mind catboxing any of your gens?
>>101960491https://www.youtube.com/watch?v=gopKHVBlJJQ
>>101959769>schizo post deletedThe rentry is working remember to follow the advice
>flash attention, xformerswith modern pytorch, relevant how?
>>101962039How about one of us sexy and handsome anons come over to your room and you know...show you the basics, help you get set up juuuust the right way, get that forge up and running so we can gen aaaaaall night together?
>>101962098PREDATOR!
>>101962098sure, seems like youre up for itdidnt know sdg fags moved over right as SAI got btfo'd
>>101960635very cool aesthetic anon, nicely done
>>101962113sdg was never usable honestly. Even /b/'s general, for how much they ignore normal convo or questions was more useful.
My boss at work is annoyed with me again, I should just quit this job, she's so annoying always telling me how I can't do basic things like printing (the printer is so confusing)
>>101962127exactly, ldg is better, more open and helpful.. at least I thought so until someone has to ruin it for all like how they ruined sdg with all the nonsensical drama everyday
>>101962084I made new installs of forge and comfy for flux. For them i don't specify anything and let them run whatever it chooses. On older installs for sdxl and sd1.5 i explicitly set xformers since it's what i've used and i think it was marginally faster than sdp, or whatever that was, when i tested it.
>>101962121ty friend
>>101961172Oh noes, it's got a signature. This must be plagiarism and copyright theft. I'm sure the artist who actually painted that would be upset to see how it was stolen by you.
someone pls post the .gguf nodes for comfyuiim going in dry.
>20 seconds and 10gb vram to caption 1 imagefuuck, I thought this would be the easy part
>>101961683lol the fuck does lcm do on 100 steps?
>>101962015ok this is with t5 and clip, didnt get any ooms, but quality defintely seems better
>>101962204still better than manual
Coomer here. So I'm currently at epoch 130 of training a flux lora on ~2000 images of a NSFW concept. Holy fucking shitballs it just werks. There are a couple of subconcepts it hasn't learned well (probably because T5 doesn't understand the tag I used at all), but for the most part it is completely mogging literal months of iterative R&D I did with SDXL models on this dataset. And I'm sure I'm not training it optimally at all, I just used the default settings Kohya recommended in the readme. A good full finetune on a few tens of thousands of NSFW images, and this thing will absolutely BTFO every single SDXL porn model. We are actually so back.
>>101962252Finally some good gens from forge, thanks anons, really appreciate all the help!
>>101962339oh no no "flux doesn't do art styles" sisters we're losing!!
>>101960491Prompt? Did you write something to add the grain or is this a lora effect?
>>101960002For this tech to be at the fingertips of idiots has really bad implications for open source.
>>101962384Thinking logically about it, the only reason Elon would give them a completely uncensored model is he wants a scandal and regulation built around it.
>>101962204>>101962326I also find that Flux LoRa werks even with dataset made for PonyXL in danbooru tagging styles. The tags become a long ass activation token it seems like. Pairs that with a long winded english description, you can get accurate gens out the gate with old datasets. However, Flux autistically follow English prompts. Random traits that PonyXL "just knows" when you train need be mentioned in plain language. Which can be hard cause tags are just difficult to describe.
>>101962384idiots can also light themselves on fire with a lighter. also, lemme take a pencil and draw a topless harris OMG>>101962402that model is far from being completely uncensored.
>>101962384You could apply this same argument to literally any technological progress ever made."Ugg think fire is scary when used by other tribe"
>>101961214Adaptive Guidance node speeds it up significantly.
>>101962402>logic>elonchoose one
>>101962366Nah, just a slop prompt>A dramatic and atmospheric image of Donald Trump and kamala harris standing in a deep, dark cave reminiscent of the one from the movie Alien (1979). The cave should have an eerie, otherworldly ambiance with dim, unsettling lighting and a rugged, uneven surface. Surround the couple with numerous Xenomorph eggs, adding to the ominous atmosphere. They are both wearing their usual business suits, and Trump is giving one of the eggs to kamala harris with a serious face. There is a speech bubble coming from him with the text "You have to eat ALL the eggs." Kamala Harris is gazing at the egg in his hands intently, her expression reflecting a mix of fear and curiosity. The scene should capture the tension and mystery typical of the Alien franchise.You do need both Kamala and xenomorph lorasHave a kamonald trumpis
>>101962416catbox this hot mamasita please hermano
>>101962331Nice, good to know it's sorted. Sad I can't come over to help you now. Maybe next time.
>>101962463everything is publicity for xitternot one day out of the news for like 5 years now
>>101962441Well, I guess the only assurance is that everyone on X accurately says "Grok made this". Even if it's Flux, this is what Elon wants Grok to make after all.
>>101962486finally some good shit
>>101962402or he actually believes AI should be uncensored and it's going to die on the first amendment hill making it precedent that the outputs are all that matters
>>101962484all good, thanks again!
>>101962512I was so mad when he was chestpounding that it was all GROK made lmao>>101962507thanks, its been a while.. flux is a massive jump from prompting realism in pony!
>>101962430Yes, flux not learning certain tags even with lots of training is kinda annoying. I need to do more experiments, but I think that if T5 has no idea what the tag is when processed in isolation, it doesn't generate a reasonable embedding for those tokens, so the DiT struggles to learn it. I'm thinking about writing a script with a huge list of manual rules to convert the tags into basic natural language descriptions that T5 should understand better. That might work. Luckily with my dataset the set of tags is of relatively small size so it should be doable.
>Q4 gguf is 29s/it VS the original 300 or so s that i rememberhonestly, huge improvement, but i'm gonna wait another 2 more weeks kek
>>101962486what a gen BRO. solid 11/10. ima light myself on fire for the occasion
>>101962545>>101962510reply meant here
>>101962486prompt and lora?
>>101962460Thanks, i'll look into it
>>101962441>>101962442mfw anon makes the exact same point with the same example 2 seconds before me
>>101962557still trying to understand flux, really amazed by how well it understands prompts
Hi guys, total noob here. I download latest Forge and flux1-dev-bnb-nf4-v2.safetensorsFor dimensions 896x1152 I'm getting ~30s/it for GTX 1070 Ti, is it ok speed or abnormal? Could I make it any lower? btw I tried shared memory/async thing, but it was always crashing on those
>>101962597are you serious? im using a gtx 1080 and getting this on 1024x comfyui >>101962554fuck this game really is rigged from the start
>>101962583prompt was just >bath too narrow blurry FEET woman's sexo legs black and white 9mm used skypaintcumfluxdev.art website
>>101962586
>>101962615Sorry I don't fully get it. Is it good speed?
>>101962597>is it ok speed or abnormal?sounds about right
>>101962632of course its pretty decent, even compared to my gpu that should be crushing yours but im getting worse speeds on a lower resolution.
>>101962583couple of rngens to get right one, but prompt goes something like ark atmosphere, grainy, black and white photo, Hot babe in a bath tub, shot from behind, dark wet hair, photo realistic, polaroid photowith that realism lora and ancient shadows lora
>>101962637>>101962643oh dang, thanks. Yes I agree GTX 1080 should have worked better.
>>101962597seems fine. ppl come in here and mention 7 minutes per gen, so there is that.
I can never concentrate when I'm at my work meeting.
>>101962159thanks anonAccording to pytorch and huggingface there should be no reason to use xformers at least on modern torch
>>101962349It doesn't. Try doing impressionism or non realism.
>>101962084pytorch have flash attention implemented since 2.2
>>101962689Are you using a quant or fp?
>>101962717Also if you can't see the AI slop in the outputs, rather than properly stylized ones (that can only be fixed with a LoRA or finetune), that is on you.
>>101962672I see, thank you
>>101962616you lied to me
Come and get your own loaf of...>>101962774>>101962774>>101962774
>>101962739NTA but I have modern Torch installed and comfy had has that "not compiled with flash attention" anyway. Not sure why.
>>101962786It obviously doesn't work on windows.
>>101962778sorry that was the metal toenails shaving rash one
>>101962741I'm using Q8_0.gguf
>>101962786Same