Discussion of free and open source text-to-image modelsPrevious /ldg/ bread : >>101900433>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
If someone knowledgeable can pitch in on this I'd appreciate it>>101902601
Fuck good thing I always refresh before baking, almost double baked
Blessed thread of frenship
>fuck amd >google and amazon make dedicated chips>only for their cloud>arm makes only inference chip
>>101902653cursed shithole for nonces
>>101902676Go away non-fren
Noob question. Is the main (practical) difference between flux and flux-dev that the second one is less censored?
>>101902663kek
>>101902700You mean Flux Pro and Flux Dev?Flux Dev is a distilled model of Flux Pro. It is weaker and produces worse gens than Flux Pro. Pro is SaaS only, Dev is a local model. There is also Flux Schnell which is a "fast" model which gens in 4 steps at worse quality.The censorship is the same on each, and meaningless because people have already made nude LoRAs.Beyond nudity and porn, violence is also censored. Other than that it's pretty open, and there is very little you can't gen.
>>101902700Flux Pro is only available through BFL's API or replicate and fal.aithese all have some form of restriction but I haven't seen anyone post a Flux Pro gen that isn't possible on FluxDev
>>101902739>>101902741Thank you frens. Are LoRA's possible on GGUF quants? I think that's the only thing I will be able to use in the foreseeable future if at all.
>>101902741All of my gens tonight have been from fluxpro.art because I need my gpu for other tasks. I have run into 0 model restrictions that don't also apply to dev. >>101902766>Are LoRA's possible on GGUFYes, if they are possible on nf4 they are possible on gguf, all it takes is for someone to do it.
>>101902778what do you find is the biggest difference between Pro and Dev?
GGUF anon, someone is asking you to quant Kolors aswell kekhttps://www.reddit.com/r/StableDiffusion/comments/1eso216/comment/li7qhen/?utm_source=share&utm_medium=web2x&context=3
>>101902788Quality
https://imgsli.com/Mjg3ODgy/1/3Here's a comparaison between fp8/fp16/Q8_0/nf4 on a house
>>101902824GGUF Q8 is even closet to FP18 than any of them. Very impressive. Only 12GB right?
>>101902824I never expected fp8 to be so different to fp16... damn... glad that Q8_0 exists now lol
>>101902840>Very impressive. Only 12GB right?yep, just a little bit bigger than fp8 (8bpw) because Q8_0 is a 8.5bpw
>>101902848What is the gen speed compared?
>>101902855You have all the informations there:https://reddit.com/r/StableDiffusion/comments/1eso216/comparison_all_quants_we_have_so_far/
>>101902866OK but what about without cfgmaxxing?
>>101902896without the cfgmaxxing it's twice as fast
>>101902900So essentially the same as fp8, which, given the quality improvement, is insane.
>>101902906as a /lmg/ fag it's not surprising, fp8 is just a dumb quand method putting all the weights at 8bit, there's way more sophisticated methods than that and GGUF is one of them, but yeah, Q8_0 is really really close to fp16 now, that's cool, I'm using this shit now
>>101902916I really should hang out in lmg more but it's hard to think of a use case outside of sexbots which I don't care about.
What was the idea behind having Miku with dreads the standard for comparisons since flux came out.
>try loading flux q8 gguf in kpnold CPP>Crashes Wtf?
>>101902944nta but because it's a gen that you cannot in flux do without using dynamic thresholding and altering the cfg. Miku is way too powerful in the model and always takes over a gen.
>>101902944It's because I found that at cfg = 1, it was impossible to me to add dreadlocks and black skin to miku, so it basically became for me a good way to measure prompt understanding, and one way to make it work is by foing some CFGmaxxinghttps://reddit.com/r/StableDiffusion/comments/1ekgiw6/heres_a_hack_to_make_flux_better_at_prompt/
>>101902961>>101902969fix your internet nigger
>>101902975I made a typo, sowwy :3
Bros.... we are so backThe future has never looked this bright
>>101902981anon, don't be a grammar nazi on yourself
>>101902981This isn't reddit nobody cares.
its silly how good FLUX combines women and tentacles
>>101903040I'm more interested in how you made it gen a camel toe
https://github.com/city96/ComfyUI-GGUF>LoRA / Controlnet / etc are currently not supported due to the weights being quantized.>currentlyI hope that means that it's possible at the end
>>101903054Mate, 2 weeks ago it was impossible to train Flux. Look at us now. Give it 2 more weeks.
>>101903066true true, I shall never doubt the power of flux after all the good shit that happened in such a short time
i predict that next month we'll get a new model that destroys flux and puts local image gen at the very top of the food chain
>>101902739How about non-photoreal stuff like "3d" illustration style fantasy art, paintings, anime, etc >>101903040I'd have excellent gens of this genre from a sd1.5 model
>>101903054I mean, what happens if you train on a quantized model?
>>101903104ya me to, but I wonder what will happen with this if it gets mangled to make such topicsalso, some bad advice: dont ask flux to have a dog sitting on her, might shock you
>>101903124you don't need to train a lora on a quantized model to make it work, on the LLM ecosystem, they can use a full precision lora on top of a GGUF quant
>>101903104yes
>>101902778>>Are LoRA's possible on GGUF>Yes, if they are possible on nf4 they are possible on gguf, all it takes is for someone to do it.nf4 for flux is already done by bitsandbytes
>>101903145are these flux pro gens?
would it be possible to get a Q8 of the T5?
>>101903221it's way easier, t5 is a text encoder, the Q quants were made for text models
>>101903227that's nice to know, because the fp8 t5 is garbage
>>101903191Yes, dev can do these too though. picrel is dev, and 99% of the gens you see here are dev.
>>101902957>>try loading flux q8 gguf in kpnold CPP>>Crashes >Wtf?I need this too. Or something that allows me to share my two 12gb 3060's for loading the model instead of using the other for t5, clip and vae.
>>101903234i guessed these were flux pro because they are way more "lively" than your usual flux dev gen >>101902812is flux on grok the pro version?
>>101903234reminds me of of her
>>101903247They were basic gens, and very doable on dev. Literally nothing special about them. The prompt was simply>"style" fantasy art
>>101903247you can nearly anything with flux (but genitalia) do your work and prompt correctly
>>101903265>>101903298Definitely, they are all doable, i never questioned that. just the composition and lighting are way weaker, and most gens look "muted" by comparison
How mad is SAI right now that their dream of landing a contract with THE social networking site was just taken from them by BFL?
>>101903364your prompt is not generating this consistently
>>101903332Considering they weren't even in the running, probably blindsided. Midjourney was close to being the model used but allegedly they backed out last second.
Guys, what if I told you I managed to remove the blur on photos with the ToneMap node?https://imgsli.com/Mjg3OTE5https://github.com/comfyanonymous/ComfyUI_experiments
>>101903054>https://github.com/city96/ComfyUI-GGUFDoes GGUF give better performance on CPU/APU?I'm wondering if it'd support Vulkan?
>>101903364You could have just posted a catbox so I didn't have to see this
>>101903385I think SAI should suck a fat bag of dicks for almost getting away with intentionally killing image diffusion for reasons I still cannot comprehend.
Is 1.5 dead? I feel left out with my shitty GPU.
>>101903409Yes, but how shitty is your GPU?
Why flux generation on this anons macbook look so good compared to the flux generation hugface?
>>101903418AMD's RX5700XT
>>101903420
>>101903426
>>101903409Flux can gen ok images at low res.
>>101903409no 1.5 is still very valid it can create more with less prompting>>101903426thats kinda sadbut the idea is .. FLUX is so mighty you have to tell it everything you want to see, while SD15 or SDXL just make the imagination spin of a few simple words
>>101903433vs https://huggingface.co/spaces/black-forest-labs/FLUX.1-schnell
>>101903444
Why the fuck is image generation at Q8 gguf at 1344x768 resolution taking up 23gb of vram?
>>101903442Thanks for the cope. I will defer killing myself for a few months and see if rocm 6.2 let's me run flux on this card.
>>101903472Because you are unzipping the file
>>101903442>but the idea is .. FLUX is so mighty you have to tell it everything you want to seeDataset and tagging, dalle works just fine with simple prompts and is more creative. Nothing that a good 2M+ images finetune can't fix. I wonder which model is going to be the next lightning in a bottle.
hmmm, i am trying the new flux gguf quants.is it supposed to be this slow?also feels like the q4 quant uses way too much memory.
>>101903482wrong thread anon
>>101903496>I wonder which model is going to be the next lightning in a bottle.One that hasn't started being trained yet
>>101903497the workflow i use.
>>101903493That is not how ggufs work.>>101903482I think you're in the wrong thread, but those are the raw weights. You either want an EXL2 or a GGUF version of the quant your computer can handle.
>>101903497no it's not supposed to be this slow
>>101903508pony was what? 6 months after XL? i think we need a bigger timeframe for this model.
i'm looking for something like the Power Lora Loader but for text prompts instead. i want to be able to toggle prompts on and off and add them to positives or negatives, instead of having to write them and delete them all the time. does this exist?
>>101903510I don't know why it's slow but negative prompts don't work so delete that, and prompt weights don't work so remove those too.
>>101903393What happens to the blur if you change your scheduler to sgm_uniform and the sampler to dpmpp_2m?
>>101903524>i think we need a bigger timeframe for this model.It took fucking ages before good LoRAs started coming out for XL too, and we have a shit load of them in 2 weeks. Flux is exponential. I expect fine tunes much sooner.
>>101903409I am not interested in anything I can't train myself locally. with 12GB, I am stuck with 1.5, but I am happy with the results.
>>101902631the future is here
Is it normal that vram usage doubles when using cfg with flux?
>>101903595yes. also leave cfg at 1 unless you are using dynamic thresholding.
>>101903595like you jump from 8gb to 16? no it's not normal at all
>>101903510You can bring your image width and height down too. Then if you see something you like you can save the seed and rerender at a higher resolution
>>101903541yep, still working anon
what sampler for flux
>>101903618No, I mean if you disable the tonemap
What are the chances that this grok shit gets fluxed nerfed before it even gets started? Elon is using twitter to bring flux to normies, they are already genning taytor swift lewds and lawsuits are likely coming.
>>101903496>dalle works just fine with simple prompts and is more creativeyour prompts get rewritten by GPT before being fed to DALL-E 3
>>101903629I can't disable tonemap, cfg = 6 would burn the picture
>>101903619euler+betaheunpp+beta if you have the timeddim+ddim_scheduler if you are going high CFG
Anyone got amd to work with zluda and comfyui and flux on windows?
>>101903634on bing, not at all, on gpt+ you can ask gpt to not rewrite it, though you need a soft jailbreak for it to stick.
>gguf>still have to load CLIP separatelywhat is the fucking point
>>101903651>on bing, not at allunless that's a new development then yes, especially on Bing, your prompt gets rewritten
>>101903465(American political stadium republican rally background), [(Taylor Swift with goth makeup), polaroid photo with VSCO filter, 1990, at night, flash photo, selfie, detailed, beautiful shadows, DSLR]
>>101903665>he doesn't knowunfortunate
>>101903663that would be retarded to not separate them, it's always the same clip, I don't want to download the same clip everytime I wanna try a new flux model, I'll just download the unet model
I understand this kind of thing happens when a model lacks training data on a particular concept, or on low resolutions (it has to spread attention over a lot of things in the image and does all of them bad, even if it could generate a single image focusing on each and do well).What strategies do you use to correct this? For simple cases, upscaling (hires fix) or genning at a higher resolution usually does the trick. Other than that I can only think inpainting each portion of the image to fix it individually.I'm a comfyui noob, so maybe there are comfy-specific strategies, too.
>>101903663why is that an issue? you don't want CLIP and T5 tied to the diffusion model, it would be a waste of space and bandwidth
>>101903643hmm in my tests on xl sgm uniform was a standout for scheduler but i will try that
>>101903679Upscale to schizo res, face detailer, then downscale
>>101903671Nice prompt.
>>101903671what's with the brackets
>>101903603Yeah cause when I use the q8 gguf with cfg, it's using like ALL my vram. It's normal with no cfg, around 13gb
>>101903758that's weird, I don't have this issue at all... maybe you should make a detailled issue on his Github so that he fixes ithttps://github.com/city96/ComfyUI-GGUF
>CFG is a must for expression>>101903758
>>1019037021.5 has a special kind of soul
>>101903702(American political stadium rally background), [(Taylor Swift) blue eyes goth makeup showing off her cute feet, masterpiece, (photorealistic:1.4), best quality, soft lighting, photograph of a beautiful [European|White] young woman with long blonde hair, RAW photo, 8k uhd, film grain, (fisheye lens:0.9), (cowboy shot:0.9), (bokeh:0.9):2]>>101903720It adds the character on top of the background, but comfyui has that feature already so not necessary
>>101903795Damn this looks realistic, with some editing would seem like something she posted here
>>101903827>Damn this looks realisticnigga what
>>101903803pic related>>101903827It's epicphotogasm (sd 1.5)
>>101903803kek, feetfag's wet dream
>>101903516>>101903614>>101903533thanks for the input, much appreciated.changed the prompt and did 512*512. still slow but better3/40 [00:27<05:41, 9.24s/itah well, i'm sure the nerds will make it retard prove soon for people like me.once its done it does output a good image though.wish i was younger with more time on my hands. had to beg artfags for rpgmaker assets. and i remember the 64*64 horror images we had a couple years ago. really cool stuff.
>>101903774Yeah, I'm gonna post my workflow here and see if anyone wants to see if they get the same thing happening. Maybe I'm not seeing something in the workflow that's causing it to eat twice the vram it should be.https://files.catbox.moe/ws73hx.pngIf it's not just me I'll make a report to the github
>>101903868Make an issue on his github, maybe he'll fix it: https://github.com/city96/ComfyUI-GGUF
Can we now close the debate that FLUX cant do picasso Miku?
>>101903889only if you catbox it
>>101903889>No messi>No obscure sport from final fantasy that that retard claimed was a common piece of well known trivia that all models should knowI'm still not convinced.
>>101903671>>101903795>>101903842Nice, this is what I get
>>101903924and this
>>101903932there's a TS lora for Flux Dev
>>101903889We closed it a week ago when we were posting picasso mikus
>>101903889okay https://files.catbox.moe/xay1ip.png>>101903918messi? explain
>>101903393I guess dev only?
>>101903924I was feeding the same prompt to positive and negative conditioning... I'm so fucking stupid.I think I'm going to start saving for a new card. Iterating on this piece of crap is a pain in the ass. Not to mention I'm stuck on 1.5 forever.
>>101903965you can try it out on Schnell, the most important parts are the ToneMap node (multiplier 0.2) + CFG 6 + GuidanceNeg 10
>>101903963>messi? explainSome guy was claiming the model was a failure because it couldn't do Messi and some very specific from a final fantasy game. When people called him a retard he got very defensive.
>>101903968That doesn't look too bad for 1.5. Background starts getting messy, but foreground seems ok.
>>101903968The other ones looked better>>101903932>>101903924>>101903842This looks most realistic>>101903859What settings are you using exactly?
flux finetunes already coming in, nice
>>101903995The original 512x512 looks better (and worse, at the same time). Bilinear scaling seems to help with preventing the denoise from the second pass from messing up the details it got right in the first one.>>101904007>The other ones looked better>What settings are you using exactly?comfy defaults but I accidentally used the same prompt for positive and negative. Model is epicphotogasm last unicorn.
>>1019034091.5 forever, 1.5 will never die
>>101904020Sampling method and steps?
>>101904009But is it gguf or is it nf4, is it dev or is it schnell does it loras work with comfy or does it need to be recombobulated?
Does comfy have consistent ways of getting around the fucked up faces when the face takes up a smaller area of the picture, that isn't just img2img?
>>101903994just make a lora out of messi>>101904020>>101903968>>101903932>>101903924
>>101904033euler with normal scheduler, 8 cfg, 20 steps
>>101903510to make sure it's not leaking to ram try offload clip to cpu device
>>101904042>(cubist:1.15) portrait, taylor swift
>>101904034dont care i have 24GB lol
>>101904042>just make a lora out of messiLiterally everyone was telling him that but he kept doubling down.
>>101904045This is what I get
Thought Flux would be another SD3 so ignored it since it's been out.That was obviously a mistake.Now i'm at the beginning of knowing, again, reee.
>>101904079Mmh... Want to post full workflow so we can test properly?I'm generating at 512x512 btw
>>101903963Sonic is too powerful. I am positive Sonic has the strongest style in the model.
>>101904079>(American political stadium republican rally background), [(Taylor Swift with goth makeup), polaroid photo with VSCO filter, 1990, at night, flash photo, selfie, detailed, beautiful shadows, DSLR]euler/normal8 cfg 20 steps
>>101904099I'm using the same settings, 512x512, 20 steps, euler, epicphoto. But on sd.next not comfyui
>>101903981Do I have to let negative prompt empty or put bokeh, blurry etc?
>>101904106I will try it
>>101904123Strange
>>101904132look at the picture again you get all the answers>>101903393
Queen Elizabeth looks like a Monty Python skit.Also I need to train a Monty Python LoRA now.
>>101904106I guess you are wrong
>>101904170I only did one gen. What did you change about it?
I love city
>>101904170baker you better put this in the collage holy shit
>>101904150Are you using comfyui? On linux or windows?
>>101904189making it taller helped
>>101904189not much I just replaced Hatsune Miku/Taylor Swift with Sonic the Hedgehog with this exact workflow >>101903963
sdui has this way of inpainting where it takes a square image that includes only the masked part (instead of the whole picture), and then patches it back into the original image when it's finished. Is there a way to do this with comfy?
https://github.com/intel/AI-Playgroundintel arc bros..
>>101904266Also, how do I tell comfy if I want to use the image colors, random noise, or the existing image underneath the mask as a starting point for the inpaint?
Does any anon have a good ComfyUI workflow for Flux? I'm on a 3090 with 24GB VRAM, can I speed it up somehow? 30-45s per imagen is cringeRunning the one in here https://comfyanonymous.github.io/ComfyUI_examples/flux/
>>101903100Bro, we can barely run flux. How about we just first wait for nvidia AI accelerator monopoly to end before we dream about something better.
>>101904316look here>>101903963
>>101904316In the default workflow set your weight_dtype to e4. I gen in 20s on a 4080
>>101904266>>101904294Nevermind, found crop and stitch. I have to get used to check for custom nodes that do the things I want to do.>>101904231Comfy on linux. But I'm no longer using the same prompt for positive and negative (obv)
>>101904343>wait for nvidia monopoly to endYou seem to believe the free market is real
>>101904315fix the lighting on ronald and it's perfect
>>101904045Goddamn nigger! 8 cfg? Are you trying to prove a point?
>>101904115Very consistent on that dick sucking lips. Hair is way better in this one. Chest is a little weird, but pretty good!
>>1019034091.5 is still the most aesthetic I've found, specifically the Noosphere model mixed with Azovya
>>101904303Damn. For a cartoon that shit has more kinematic and physio-biological structure than almost anything I have seen.
>>101904448How the hell are you getting cute Taylor cameltoes? What prompt magic is this?
>>101904577what do you mean? can't you prompt?
>>101904561CFG 8 is perfectly normal for 1.5 models, retard
I'm using Fooocus and I've got this LORA with a slider but I don't see how to actually make the slider be in the interface, what do?
>>101904619No. I am a promptcel that can only get the minimum images rendered. I would like to subscribe to your news letter.
>>101904636I apologize, I haven't used the model for the poors, only flux.
>>101904616>cutelook again lenny
>>101904638?You mean the strength of the lora? sliders usually work by changing the strength>>101904655outing yourself as a newfag, there was SD 1.x to use for a long while
>>101904655>admitting to being a newfag riding on the bandwagon of Fluxnot very good at this, are you anon
>>101903409don't feel left out, sd15 can do magnificent things.>>101903842this one is scary good
>>101904667>>101904668I am a newfag. That is way came here to learn from oldfags that were making tranime out of pixels before generative models even existed. Teach me sensei!
>>101904678my sides!you SIR have TRULY won the internet on this blessed day!!
>>101904644just type what you want.. write it out like and essay and you get what you want with flux (anything but porn)
Do you clean your dataset tags? For a long time I have simply removed low entropy tags, i.e, tags with really high or low count.I've been trying to find a way to cluster images based on tags so I can automatically remove outlier images.
>>101904690Not mine
>>101903671
>>101904718oh my sweet summer child
>>101904724sir, pls help me. I'm doing it for my daughter
https://reddit.com/r/StableDiffusion/comments/1estj69/remove_the_blur_on_photos_with_tonemap_an/I highly suggest you to try this new clip_l thing, it really improves the quality picture and prompt adherancehttps://huggingface.co/zer0int/CLIP-GmP-ViT-L-14/blob/main/ViT-L-14-BEST-smooth-GmP-ft.safetensorshttps://imgsli.com/Mjg3OTU5
>>101904739seen it, noted, thanks
>do the needful
>>101904231bro you're using a LCM model. You need a lcm lora and cfg of 1.5-3. Look it up on civitai or just use a normal model for now
>>101904738remove irrelevant tagsconsolidate important tags (e.g. if you have several tags for the same feature, pick one and use that for every image where it appears)be consistentif you have extra time add missing image composition tags since afaik most autotaggers don't apply them
>>101904761If it's time gated i don't really care.If it's "premium loras", it's just retarded.
>>101904738also the most important thing for datasets is to not have any shit images, GIGO is still fundamental computing principle
>>101904720How>>101904767I'm not using any LORA, what do you recommend?
>>101904616I'm using SD 1.5 (epicphotogasm last unicorn). If you look closely, the image is a mess. I had to inpaint the face, too.
>>101904739thanks for figuring out good baseline numbers, at this point i've just started coming to /ldg/ to see what you're working on next lmao
>>101904739is there a TE only of the BEST version?
Kino Sd 1.4
>>101904790https://files.catbox.moe/el38ck.pngi don't have whatever comfy feature anon had just slapped the prompt in
>>101904829can you elaborate? >>101904819aww, thanks for the praise anon :3
>>101904851WTF are you using
>>101904865ViT-L-14-BEST-smooth-GmP-ft is the only thing, text encoder and image encoder, the text encoder is only 1/3 the size like ViT-L-14-GmP-ft-TE-only-HF-format
>>101904790search lcm on civitai. "Theally" has one for 1.5. But LCM/Turbo/Lightning models are made for speed so the results are usually not as good and they are way more restrictive (you need the lora + cfg 1.5-3) so again I recommend you just use a normal model that doesn't have LCM/Turbo/Lightning in its name
>>101904851Very nice. Care to share your specs?>>101904899https://www.nayuki.io/page/png-file-chunk-inspector
>razzz making Flux celeb loras>they are shitEvery time.
Is there a way in comfy to have it automatically start with "extra options" and "view queue" ticked?
>>101904916Yeah, it's the new shiny clip_l, you should use that one imo, Flux has all of these (unet + clip_l + t5xxl + VAE), that huggingface is only about a modified clip_l
>>101904916*is the whole thing>>101904967anon that's not... nevermind
>>101904899>open up reforge page>go to issues>someone is requesting flux support>read comment saying to use a certain branch if you want to use flux>google how to install just one branch of a repo in git>install that shit>spend like 10 minutes testing some baseline settings>go on /g/>post>>1019049323060 12gb
>>101905006forge is back up and running fella
>>101904983I think I start to understand what you mean by that, the original clip_l is 256 mbhttps://huggingface.co/comfyanonymous/flux_text_encoders/blob/main/clip_l.safetensorsAnd this one is 932 mbhttps://huggingface.co/zer0int/CLIP-GmP-ViT-L-14/blob/main/ViT-L-14-BEST-smooth-GmP-ft.safetensorsI guess that it's a bigger version of clip_l or something, desu I won't cry for a 700mb increase in my vram if at the end I get better prompt understanding: https://imgsli.com/Mjg3OTU5
>>101905032yeah and then he's gonna disappear again in a month
>>101905036it's not bigger, CLIP is a text encoder and image encoder, we only use the text encoder part.90% sure Comfy knows to load just the text encoder if you use the file that has both parts
>>101905055Tbh I didn't notice an increase in vram in the text encoder part, it's still 9.3 gb total, so I guess that comfy is only using the text encoder part of ViT-L-smooth-GmP, which is a good thing
>>101904917This is what I get with a LORA added on sd.next
>>101905043this>>101905056VROOOM
walking to school, by Sam Toft, by Serge Najar, by Richard Scarry, Studio Ghibli (minimalism:1.9)FLUX.1 [schnell]
>>101904739hey anon, you know your code changesare all 3 files required - or just sampler_tonemap.py? since that's the node you use.not sure if they're dependent on each other or you just changed it across files for consistency sake
>>101905086>not sure if they're dependent on each other or you just changed it across files for consistency sakeit's for consistency sake, so that if you want to try the other nodes, it'll work aswell
>>101903007ok this is epic
>>101905081
it's nice that flux can do generic cute anime girl sometimes
>>101905129boink, nice one>>101905085I don't think raising the strength of a token does anything in flux. those gens are very nice.
>>101904688People will teach you when you stop typing like a faggot. Also if you see the CP poster just report and hide, don't engage.
flux unchained
trying to gguf, what did I do wrong?
>>101905081>>101905123lol not just any lora, a LCM lora. Here: https://civitai.com/models/195519?modelVersionId=424706And again: 1.5 cfg, 3-8 steps and its just better to use a different model
>>101905221DID YOU PULL?!
>>101905261Yes, Comfy is up to date via manager. Installed the node. Installed the dependency. I will try a manual pull.
>>101903444damn, republican women look like THIS?
>>101905330Nope. Maybe my workflow is just shit. Can someone please catbox a gguf workflow for me?
>>101904739So it does replace the second "clip_l" in picrel?Or both of them?
>>101905411yeah, it just replaces the second clip_l, I should include that in the tutorial aswell
>>101905417Alright thanks.What does the first one does vs the second one?I didn't find much explanations outside of "here is how to use flux".
>>101905411but there is only one clip_l in that image
>>101905433t5xxl is great at natural language, clip_l is good at tags, so it's supposed to get the best of both worlds
>>101905441NTA. Is there an equivalent for 1.5? Getting the models to be better at understanding the prompt by changing the text encoder? Sorry if this is a stupid question.
>>101902778The gguf node doesn't rely on the forge nf4 custom node thing implementation right?
>>101905434*the second one named "clip_l">>101905441Alright makes sense.Would that mean that finding a superior version of t5xxl would be even better for natural language friendly models like flux?
>>101905456I suppose? Dunno what does 1.5 uses as a text encoder, but if there's another one with the same architecture that has been improved, maybe you should try to replace that yeah
>>101905242No change with any option, starting to think the issue is with sd.next
>>101905460>Would that mean that finding a superior version of t5xxl would be even better for natural language friendly models like flux?better prompt understanding basically
>>101905469it uses the same encoder
>>101903868What is this prompt?
>>101905477dude get off this LCM checkpoint and get the regular one.
>>101905485then it means that you could change its clip_l with this one, try it out and let us know if you notice some better results
>>101905495I told him that twice already kek. Probably still uses the wrong settings
https://www.reddit.com/r/StableDiffusion/comments/1eslcg0/comment/li8myda/?utm_source=share&utm_medium=web2x&context=3Looks like the GGUF loader is now working on Forge aswell?
>>101905485>>101905497Will try and report back (in a few hours). Thanks.
>>101905515>>101905495What's the regular checkpoint? I used same settings as the guy with comfyui but in sd.next
Is there a way in comfy to have multiple "desktops", aka I'd like one purely for upscaling, one for generation, one for image to text...Next best thing would be allowing multi instance.
>>101903868(full body portrait, standing pose), (best quality:0.8) perfect anime illustration,on a black surface, on a plain black background, a pretty, happy man with short curly brown hair, wearing a hoodiethen i use InspyrenetRembg to remove the background. works pretty well.the part after background was actually randomized by an llm. i play around with auto making character cards and picture creation with various expressions.
>>101905541NTA but that looks similar to what I get when I upscale with a denoise parameter that is too low.
>>101905549ah, i fucked up.was meant for >>101905493
>>101905544Not sure what you mean... it's a web front end. Just open another tab and put a different workflow in
>>101905558Wouldn't that mess my currently open project?Basically having multiple tabs would work and everything would be separated?I'm a bit of a techlet for this.
>>101905541you shouldnt use more than 8 steps with a LCM model. And the size shouldnt go beyond 768 for 1.5. So use 512x768 for example. And you still seem to use a random lora not he lcm one. The name of that lora is LCM_LoRA_Weights_SD15. Also if you really want to download epicPhotoGasm then just download the original https://civitai.com/models/132632/epicphotogasm not a LCM version. Then you dont need to do all this stuff
>>101904739So much has improved this day>DynamicThreshold got a better sucessor, Tonemap>A better version of clip_l can replace the old clip_l on flux>Q8_0 is closer to fp16 than fp8 and can be used without much drawback (it's a bit bigger than fp8)Now with all those changes I have way more ease to make cool prompts on flux, that's nice
>>101905541https://civitai.com/models/132632?modelVersionId=223670>>101905515those underwater gens look realy good. did you also do those squids in deep water?
boys is flux just another model I can pop into my stablediffusion folder and start using or is there more i need to do to get started
>>101902289>Is it possible to use any version of flux on an RX 5700 XTOf course. Simply recompile the entire ROCm stack all the way through MIOpen and composable_kernel. And patch MIOpen and composable_kernel so they work less shitty on the 5700. Then compile PyTorch.
>>101905588Each tab is its own instance of the front end. The back end is like a "server", it just handles api calls (queued prompts) from any instance of the front end.
>>101905632Nice, thanks!
>>101905611no. Only jellyfish
Managed to run flux gguf, but not with this clip-vit-large-patch14.bin in the clip loader. I used instead a clip_l.safetensors. I have no idea if this is how I'm supposed to use it, but it worked. I found a model "clip-vit-large-patch14.bin", but it's clip vision, and when I tried to load it, the loader complained, so I guess it's the wrong model but with the same name?Anyway, I'm getting a 4.3s/it on a 4060Ti.
anyone know why this random pony model has the term "flux" in its version?https://civitai.com/models/592983?modelVersionId=724618what the hell does it have to do with flux?
>>101905651>Managed to run flux gguf, but not with this clip-vit-large-patch14.bin in the clip loader. I used instead a clip_l.safetensors.go for that one anon, it's a better version than the classic clip_l >>101904739
>>101902766>Are LoRA's possible on GGUF quants? I think that's the only thing I will be able to use in the foreseeable future if at all.Why not a node that allows you to load a LoRA separately on its own into VRAM instead of merging with model weights? The math should let you keep it side-by-side no? You'd just lose a bit of VRAM on the LoRA itself and a tiny loss of speed from the LoRA math each gen.
>>101905660Yes, the one making it clicked wrong.
>>101905627I actually got started down that path at some point. But I'll rather wait for rocm 6.2. It's supposed to work with gfx1010. Then I will be able to use the shittiest gguf.
>>101905611>>101905604Same thing, must be some setting in sd.next
Ready to go with some fresh bread...>>101905755>>101905755>>101905755
>>101905737cfg should be 1.5 not 6
>>101905737now remove that goddamn LCM lora, raise steps to 20 ish, pick a better sampler and off you go
>>101905700rocm 6.2 is out, but checking composable_kernel at least gfx1010 still isn't a default build target, so unless your distro loves you (debian) it still needs recompilation.I think 6.2 has the bugfix for rocblas so stuff like llama.cpp that only uses rocblas will just work on gfx1010, but not pytorch.
>>101905779nvm this isnt the LCM model anymore. Disregard what I said
>>101905737lora weight to 0.5, cfg to 1, steps to 6t. genned with LCM for 5 months, switched to HyperSD which let me gen in 4 steps with better quality
>>101905660Just a jeet trying to get more eyes on his shitty merge.>>101905682There was no misclick.
>>101905737
Do any of you have any idea what was used to gen this?Original PNG: https://avatars.charhub.io/avatars/similar_reply_6303/student-vs-teacher-7408d42bf204/chara_card_v2.png
>>101903007Well...I guess this is why I am here forever.
