Discussion of free and open source text-to-image modelsPrevious /ldg/ bread : >>101914501>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
>leave PC and have it autogenning>come back>turns out one of the first gens was the best by a very large marginIt keeps happening...
>no collageThread ruined already.
Purple goes great with green.
>>101917935I know that feel anon
This gives me a Dr. Stone vibe>Flux is quite good
>no collageI'm going to kill myself
This fortnite one is spot on.
We need a complete and total shut down of "photorealistic" slop until we figure out what the hell is going on.
>>101918017>We need a complete and total shut down of "photorealistic" slop until we figure out what the hell is going on.you can remove the slop with this methodhttps://reddit.com/r/StableDiffusion/comments/1enm9og/discovered_by_accident_a_trick_to_make_flux/
>>101918055
A Japanese girl in a bikini at the Beach. It's raining a little bit and the waves are slightly choppy, but you can see the sun rays in the distance.
>>1019180712/2
so much for black butler anime prompt
>>101918137its the 5gs that cause viruses, duh
1/2
>>1019181372 cents have been credited to your Pfizer food stamps account.
>>101918137>eat horse paste insteadThat line sounded exactly like that line on idiocracy
2/2
neat, reactor also works with flux in forgebehold, gosling trump: txt2img btw, probably even better swaps in img2img
>>101918225at least the horse paste would kill any parasites you might have from undercooked food
>>101918137oh no... debo has hijacked /ldg/, it's over...
>>101918233
>You find an unmarked crate of H200s unattended, no cameras around, factory sealed and large enough that you can carry in your arms.What do you do?
>>101918242I bless thee
>>101916458what is this extension?
sdxl/pony gens can also be used with forge/comfy inpainting for text or whatever you like too, fun stuff.>latent noise + 1 denoise>get signs
>>101917989derriciousi wish real asian women had asses like that
this shit should be illegal
>>101918137>bill gates and fauci want to put tiny bruce lees in your body to fight covids. its anti american. eat horse paste instead
>>101918399fuck off tripfag
>>101918366They do anon, you just won't get to meet beautiful Asian women like that easily.
>>101918431it will be illegal after kamala is electedeven though her campaign is using the tech to fake crowds.
>>101918447it will be illegal 4 u, not for them
>>101918294The people will be getting the ultimate flux finetune(s) for realistic porn and hentai, for free.
so reactor works 100% fine with flux (in forge, I assume comfy is the same), there is your loophole for celebs not in the dataset.
wow! celebs!!
>>101918294I briefly consider stealing them before realizing I don't have all the other required equipment for actually mounting, running and cooling them, and walk away
>>101918460
>>101918294A shit ton of XY plots to find the best settings for my workflow kek
guys there's a loophole, for the celebs!!!
>inpainting in comfyui still intolerable after all this time
>>101918447I certainly do hope America makes it illegal so China can be the one to develop AGI.
I think I'm done with heavy image genning for now. Time to take a break and go back to text gen for a while. Miku could use a break from all the walking herself.Good night /ldg/.
>>101918475They cost for the peripherals would be miniscule compared to the profit you could generate from any number of uses for a a crate full of enterprise GPUs.Play your cards right and you have a legitimate passive income from the rental alone.
>>101918499it's a shame, I just keep a Fooocus install around for inpainting but I wish I didn't have to
corporate miku:
what the FUCK is bitsandbytes and how do you install it?
>>101918532pip install bitsandbytes
>intelligent models have a slopped style>soulful models are retarded and don't understand promptswhat do
>>101918539in which folder? main?
>>101918548don't think it matters
>>101918447Flux is open source. fuck the police. You can't stop Flux from existing on every computer. The results of the election won't stop the existence of open source AI. It will continue to spread and can not be stopped. No amount of zionist influence can control open source AI. It is the future. It is inevitable.
>>101918548Open the file explorer, go to the comfy ui custom nodes folder and the nf4 (which I assume you're using) and type cmd in the top barthen write, pip install bitsandbytes
>>101918562>Flux is open sourceWhy does everyone keep saying that? It's not open source, they just threw us the weights to play with.
>>101918548do not use nf4 dude, it's worse than Q4_0 (which has the same size)
>>101918435>They do anonnopeI've never seen an amazing-assed azn until this thread and I'm almost 40
>>101918557>>101918566did it in main, I don't have any nf4 folder yet because it says to first install bitsandbytes and then the node>>101918579this shit is esoteric as fuck, the only way to know for sure is to test all this crap for yourself and compare
>Trump is pro crypto and pro Artificial Intelligence>Kamala is anti crypto and and anti AIwtf I love Trump nowshit, I wasn't going to vote because the system is a joke and I don't trust the government but now this might convert me into a voter again
>>101918597>this shit is esoteric as fuck, the only way to know for sure is to test all this crap for yourself and compareThe simple fact that nf4 is the only quant who decided to have Miku as a different pose than the rest is a huge redflag, this shit stinks
>>101918598well duh, democrats are pro censorship and republicans are pro freedom, nothing new in the sun
>>101918611I did test it on forge and it kinda suck BUT I have everything in forge sucks by defaultif I can install this fucking shit I'll report back
>>101918611He's right. They should all produce nearly the same image but with spot the difference levels of variation. Nigu being in a different pose is concerning.
>>101918625i love pro-freedom politicians that require me to swear oaths to israel if i want a government job
>>101918627forge supports both nf4 and gguf quants, so you better go for Q4_0, it's not really good but way better than this piece of nf4-v2 shit
How to make the text more clear?
>>101918640both side do that though? it's not like voting democrat will help you on that shit
>>101918640the real path is to use AI to free these polticians from Israel control
>>101918641where can I get Q4? what are the advantages? does it run in chackpoints or unet?
>>101918640both parties and candidates are pro israel so you're fucked there
>>101917989brap
>>101918673this
>>101918673yeah this is real reason i didn't want to vote. unless they have the balls to denounce Israel and have a pro nuclear energy stance in their campaign, they aren't good enough.
>>101918665>where can I get Q4?https://huggingface.co/city96/FLUX.1-dev-gguf>what are the advantages?like I said, it's the same size as nf4 but produces better quality overall >>101918579>does it run in chackpoints or unet?in comfyUi it's a unet loader because the huggingfaces gives you unet models (I prefer like this, it's retarded to have to download the same vae + text encoder for each unet changes)https://github.com/city96/ComfyUI-GGUF
>>101918374Love this, catbox?
>>101918705you're a gentleman and a scholarI'm going to test both because autism, and compare all the other models as well
>>101918705NTA but was a LoRA workaround figured out for ggufs yet? It's very high on my importance list.
>>101918723>you're a gentleman and a scholarthanks, and have fun with those anon o/
>>101918579Is it just me or are all the Q models half the speed of the F ones?
>>101918753not at all, look at the speed again, it's equivalent, only Q5_0 is fucked on that regard (my theory is that Q5 isn't a multiple of 2 so the GPU doesn't like it)
>>101918715Thankshttps://files.catbox.moe/qpqk8g.png
>>101918741>6.79 GBoh shit, so smallok, so based on your comparison table, Q4 would be perfect for 12Gb vRAM, right?the rest are obviously way better, but go beyond 12Gb so they should be considerably slower, am I in the ballpark?
>>101918753No, but using CFG for some reason eats up all my vram even at q8
>>101918779Q5_0 is a little to big for 12gb vram but Q4_1 is exactly what you need, it's between Q4_0 (too small) and Q5_0 (too big)https://huggingface.co/city96/FLUX.1-dev-gguf/tree/main
>>101918765>>101918781Not talking about the picture or cfg I mean in practice on my 4080 in a standard workflow
so I gave this a shot to compare to joy caption, https://huggingface.co/qresearch/llama-3-vision-alpha got it working with Llama-3.1-8B-Lexi-Uncensored-V2 instead of the 4bit version of llama-3-8b-Instructseems to actually be uncensored, but main issue is it squeezes the images to 384 x 384, but joy caption i dont know what it is resizing to, it handles images differnetly
>>101918809Q quants aren't supposed to be slower than the fp ones if you have enough vram to run, if you don't, then I don't know how it behaves
>>101918543Learn to prompt each the way it’s supposed to be prompted. Learning is a process that involves being wrong repeatedly. Some people are incapable of enduring this discomfort and thus never learn anything. They pretend to know everything and scurry through life hiding from themselves.Now, how the fuck is it possible for AUTOMATIC1111 and Comfy to gen normally (SD 1.5) and Forge to take while minutes to for 2 measly low res iterations?
>>101918797yeah, if anything it might be worth testing Q4 and Q8I'll do so
gen made in sdxl/pony, inpaint edits (to test) with flux/forge. flux doesn't have all the character loras yet, but you can use existing loras/models together with flux, to do whatever you like. flux inpainting even without controlnets is good. (school uniform before edits)flux excels with text and many other things, but for specific characters sdxl/pony/1.5 are still the best option unless flux has the character in it (like Miku)
>>101918353>I open an image editor>I type text onto the sign???
>>101918435That ass is as fake as those “Japanese palm trees”. Clearly a Californian that has never been to Asia conjured that image. A tiny Californian living inside your computer.
>>101918811joycaption certainly isn't 384, it can pick up some pretty small text
>>101918598The companies that run your country have already decided how it’s going to go down. Your candidate of choice will only affect the flavor of the whining from the plebeians.
>>101918705what about Q4.1 and Q5.1, have you tried those?
>>101918853
what does this mean? happens every time I try to run a gguf >\ComfyUI\custom_nodes\ComfyUI-GGUF\dequant.py:10: UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() or sourceTensor.clone().detach().requires_grad_(True), rather than torch.tensor(sourceTensor). data = torch.tensor(tensor.data)
>>101918502Neat gens
>>101918872she has momface (positive)
>>101918867We know how the Q behave, it's like this;Q4_0 < Q4_1 < Q5_0 < Q5_1 < Q8_0The comparaison image was made simply to see how some of those Q quants perform against nf4 and the fp4/fp8
>>101918832How the fuck is it possible for AUTOMATIC1111 and Comfy to gen normally (SD 1.5) and Forge to take while minutes to for 2 measly low res iterations?
>>101918867i cant get 4.1 to work in forge. q8 works fine.
>>101918892>>101918904roger>>101918897imo forge is fast but sucks ass, and speed over quality is useless
>>101918777Thank you bro
Finally, comfyUI can be experienced the way it was meant to be
>>101918925forgot pic
>>101918935>lightmodeliteral monster
>>101918843devil noko:surprised the inpainting is consistent in flux even without openpose or canny being used. still want all the controlnet stuff though.
>>101918935
>>101918951wonder if it's a model size thinglike how a big LLM can intuitively one shot stuff that needs to be explained to a smaller one with examplesa large image model can inpaint coherently without assistance
okay wtf is this shit?why does the search and the nodes look different now
>>101918987Because the UI was updated to look less shit
>>101918962yeah cause normally i'd need to use openpose to 100% get a figure in proper proportion if inpainting, or canny for 1:1 lineart (and it's fast, not really a problem), but flux seems to work well as is which is nice
>>101918951What are you using for inpaining? Forge was trash,
>>101919001swing and miss, I guess
only prompt: large white angel wings. anime style.made two blobs on both sides and voila, wings. I know inpainting is nothing new but flux is good at lots of stuff including text in perspective so it's also good for edits.>>101919011this is forge with the flux nf4 model, default img2img inpaint tab. seems like auto1111 to me
>>101919016>>101918987Comfy really made a UI change on a system used almost exclusively by autistics. I don't think he thought this through.
>>101919033>this is forge with the flux nf4 modeditch that shit anon and go for Q4_0 instead >>101918579
>>101919064Would if I could Anon. This is all I get.
>>101919083maybe you should make an Issue on his repo so that he can fix ithttps://github.com/city96/ComfyUI-GGUF/issues
>>101918919Nice
>>101919093Is it even a comfy issue because nobody else seems to be getting it. It's something on my side but I have no idea what
>>101919083did you update that GGUF package? he did fix some shit for the vramlets
>>101919061well I think comfy's UI is utter trash anyways, but the point is that it looks the same or slightly worse
LoRA for gguf when?
>>101919108maybe there's more people having the same issue as you and aren't as vocal about it and decided to give it up, you're different anon, you're a man of sheer will and determination, and you will make an issue so that he can look at it and make it work with you
>>101919129already possible on forge, the gguf anon said he's working on it to make it work for the comfyUi node
wtf is "distilled cfg" vs regular cfg
>>101919203distilled cfg is a cfg that works on flux because the regular cfg burns the picture at cfg > 1, if you want my opinion, distilled cfg is a fucking meme and nothing beats the old cfg, fortunately you can turn cfg > 1 without burning the image with Dynamic Thresholding https://reddit.com/r/StableDiffusion/comments/1ekgiw6/heres_a_hack_to_make_flux_better_at_prompt/
>>101919235For how much you've been shilling your process you'd think your images would look better than they do
>>101919250srsly
if i want to train a flux lora, i should use the fp16 model, right
>>101918811yeah, strangely this llama-3-vision-alpha can read text (like their book example image) but it tends to go off the rails and skip over stuff someone made a nice modificaition where you can change the image without restarting https://huggingface.co/qresearch/llama-3-vision-alpha/discussions/2
>>101919131I did it
>update comfy>math nodes no longer workMany such cases
>>101919083I get that too but it works, that being said, nf4 is faster
>>101919250Silence, Nigger. His CFG method has unsloppified many of my gens.
>>101919250They don't? https://imgsli.com/Mjg2Mzgy
>>101919235i answered my own question by looking in the previous bread but ty it doesn't make sense to call it "distilled cfg" when its conditioning guidance. i must have missed that one
Is there any way to get rid of that plastic look of flux dev right now? I tried some "realism" Loras but they don't really make much difference.
>>101919368I've found lower guidance of 2.0 or 2.5 can help sometimes. Beta scheduler too.
Comfy bros, they are making fun of us, bros. That Forge guy apparently has some superior tech, bros. What do we now? This Comfy cuck is leaving us wanting, bros. He's making a joke of us, bros. Should we start a harassing campaign against this Forge guy so he gives us his tech, bros?https://www.youtube.com/watch?v=90cz8prKfMY
>>101918579how much vram do you have anon?
>>101919432It says in the image dumbass
>>101919432I have 2 GPUs, the first one (3090, 24gb) is for the unet model, the 2nd one (3060, 12gb) is for the text encoder
Fucking love GGUK 4.@*
>>101919449lick my taint, faggot>>101919450dang, no wonder you get those speeds
>>101919273if you're using kohya sd-scripts he said it has to be that model or it wont run
>>101919464>dang, no wonder you get those speedsthe speed you're seeing on the picture is with GPU1 vs unet, you can get the same speed if you force the text encoder to go to your cpu, the text encoding will take some seconds, but after all you'll get normal speed on your gpu if that one has enough vram to run your unet modelhttps://reddit.com/r/StableDiffusion/comments/1el79h3/flux_can_be_run_on_a_multigpu_configuration/
>>101919486thanks
>>101919453Can you post (or link to) your settings? What are your gen times? Are you doing anything special to offload the model? Schnell or dev?I'd appreciate if you could share your workflow. I've been struggling to get usable times with 8 VRAM.
>>101919450how the hell did you do that? i've wanted to do that for ages with my handful of GPUs
jesus christ, I've practically re-coded app.py for joy caption to get it to run locally without any errors or issues. add that the requirements.txt is missing 15 different requirements so you have to manually pip installall because huggingface is cucked and won't give me full uptime on a gpu even if I pay for it. I hate this shit
>>101919521look at this reddit post >>101919490
>q4 is better than fp8what the fuckl
>>101919235Even when using your setup, I find it still deepfries the hell out of my gens with the CFG set to 6, especially on photos.
>>101919548Dynamic Thresholding is bad on realistic pictures, go for its alternative, Tonemap https://reddit.com/r/StableDiffusion/comments/1estj69/remove_the_blur_on_photos_with_tonemap_an/
>>101919527>add that the requirements.txt is missing 15 different requirements so you have to manually pip installI just used my ComfyUI venv and missed no deps lmao you brainlet
>>101918883I get this too, but it still runs, so I'm ignoring it for now since the loader is literally called "bootleg" you don't need to think too much about it>>101919083oh i see, it doesn't work for you. good chance you installed the gguf package but incorrectly and missed your comfyui environment. try launching your terminal of choice, installing it, and immediately running comfy in that same terminal
>>101919490interestingI'm on 12Gb so I doubt I'll make a differenceso far NF4 is the fastest, but the quality doesn't seem worth the speed against Q4what about the workflow? is it better to use the fp8 workflow, or the dev workflow?
>>101919573lmao I'm a literal retard oh well, its over nowits done, it worksrunning on --listen so I can comfily caption from my laptop
>>101919578Weirdly it ran for these 2 images, now it gives the same error
>>101919583>what about the workflow? is it better to use the fp8 workflow, or the dev workflow?can you elaborate? I don't think I understand what those mean? And if you want to use a fp8 model, go for Q8_0 instead, that one is the closest to fp16 >>101918579
What is the difference between _1 and _0 and which one do I need for 16GB VRAM?
>>101918597>this shit is esoteric as fuck, the only way to know for sure is to test all this crap for yourself and comparei have said it when SD startedI have said it when the chinese spellbook got postedI have said it when LLMs startedand I will say it again:THIS AI SHIT IS BLACK VOODOO
put me in the collage
>>101919612sounds like some kind of oom thing then. i also get inconsistent behaviour from quanted flux, i can run base flux just fine (but at 6s/it so i have to use schnell and it still takes half a minute), while q4 either gives me 2s/it or overflows into system ram and takes >60s/it. so for now i'm just doing pixel art on pony and waiting for the half-working flux configurations to be a bit more user friendly, since even after i get it semi working flux is still pretty slow for me
>>101919650Prompt:>Anonymous 08/16/24(Fri)19:52:58 No.101919650▶>put me in the collage
>>101919614here: https://comfyanonymous.github.io/ComfyUI_examples/flux/the FP8 workflow and the DEV workflow have different nodesthe FP8 workflow doesn't have the the dual clip loader
Flux widened the skill gap so much I can't fucking make non-slop gens now.>read the reddit post Nothing works and do not try to help.
Anyone got a workflow with multiGPU support and Loras?
>>101919662I am running 16GB on a 4080. The model seems to unload when I finish a gen, as my vram and ram usage goes down. It's very confusing.I am re-downloading the models to see if that helps.
>>101919677So you just want to whine? This isn't your blog post nigga
>>101919700Shut the fuck up I am furious and will continue to be until I have a heart attack or figure it out.
>>101919676If you want to use flux-dev then go for the dev workflow, simple as that, and change your model accordingly on the unet loader if you want fp8
>>101919706why won't you just say what's the problem? I can try to help, 2 brains is better than 1
>>101919431Forge has a lora support for all the quants, but it's pretty hacky and it "patches" a lora in every time you generate an image, even if you re-generate with the same loras and weights>>101919504My workflow is a bit of a mess and relies heavily on AnythingEverywhere nodes so I drew over it to help understand it a bit betterThe load times are with cfg 6 and negatives, so without them it's even faster. Using the modified tonemap node >>101919565
>>101919692I wouldn't bother with that, just learn to check a file's sha to confirm nothing changed it https://huggingface.co/city96/FLUX.1-dev-gguf/commit/860865faf9e5ddf33132b703afe11e21baeac32d it's easier than anything else you've done to get flux running. fwiw if i cared more to get flux running, my own next step would be to just try forge, it's probably more idiotproof than comfy. If you strongly prefer comfy's ui you'll probably be able to move back to it within a few days regardless, these nodes are barely 24 hours old. leech off the commons and let someone else improve it.
>>1019196318_0 would fit in your vram completely so don't bother with the rest.Also, what's the point of a 4_1 quant? It won't fit in 8gb vram, and 10+gb vram cards are much better off using 5_0 or 5_1
Is Flux vram usage not dependent on resolution being generated..? I seem to be able to gen at 512x512 fairly quickly, but it gets orders of magnitude slower at 1024x1024, but nobody else seems to be mentioning attainable resolutions as a factor in which quant to run.
>>101919711let me explain again:you can use GGUF and DEV (dev-fp8) models with t5xxl_fp8_e4m3fn or t5xxl_fp16, those are all inside the UNET folderthen you have NF4 and comfy's own dev-fp8 that go inside the checkpoint folder, and are not paired to t5xxl_fp8_e4m3fn or t5xxl_fp16, thae clip loader node is not present in that workflow
>>101919801>Is Flux vram usage not dependent on resolution being generated..?of course, like every single image model
>>101919779>what's the point of a 4_1 quantIt's for the 9gb vram cardsIt all works now after re-downloading the models, except my math nodes for the upscale apparently don't workSame workflow just with gguf and the clip_l changed to ViT-L and t5xxl changed from fp16 to fp8. I don't really get it
>>101919868clip_name1 should be for the t5, and clip_name 2 should be for clip_l or that VIT-L, you switched the both of them
>>101919801Because of how huge the models are (before all the quants and stuff that have been made), the people who could run it had enough VRAM that it didn't matter what resolution they genned at. With the smaller models being made and people with less VRAM starting to use Flux locally, I think in the next week or so we'll see more people test how resolution size impacts VRAM usage
Comfy quant lora yet?
>>101919881Seems to be working with how I set it. I will do the other way and see what happens.
>>101919881the order doesn't matter
How the fuck can Forge run so fast? How can one man be so good at optimizing code. I thought that was Comfy's shtick.
>>101919955>Comfy's shtickIs copypasta.
>>101919923>>101919903>>101919881Correct, it does not matter. I got identical outputs
>>101919955Same speeds for me except the UI is buggy as fuck and it regularly OOM's when comfy doesn't. Loras and inpainting work tho.
>>101919882That makes sense, thank you
>>101919955Because he's not a retard like A1111. The whole point of Forge was to make a gradio interface that wasn't dogshit.
>>101919729>My workflow is a bit of a mess and relies heavily on AnythingEverywhere nodes so I drew over it to help understand it a bit betterThank you, absolute legend.
Just found a comfy node implementation for JoyCaption. https://github.com/StartHua/Comfyui_CXH_joy_caption
>>101919955I would love to use forge but image quality looks like shit, no idea why
>>101919992Bro, I don't read chinese.
>>101920000Instructions in the README look fairly obvious just from the screenshots
>>101919992I don't get it, is JoyCaption as good as 4o or InternVL?
>>101920010>4oNo, but it is uncensored.>InternVLMaybe.
>>101919955Does it work with fp16?
I still don't understand why my math nodes broke. Can any Anon take a look and see if you can figure it out for me? I am absolutely lost. https://files.catbox.moe/19ct20.pngI must have changed something setting this up but I swear I didn't.
>>101919955Forge is downright unusable for me. I'll stick to A1111 and Comfy thank you.
>>101920037I think he updated it recently to work with fp16 yes.
>>101920028If it's not as good as 4o then it's not as good as InternVL as it is SOTA anon, take a lookhttps://huggingface.co/spaces/opencompass/open_vlm_leaderboard
>>101920000If only we could develop a technology that let us translate text automatically. Perhaps some day we might even get it integrated in the browser. That would be awesome.I guess you're out of luck until then, bro.
>>101920058And which of those models are local, Anon?
>>101920058>76BJoy Caption uses llama 3.1 8B so I don't think you should be comparing them
>>101920075>>101920058you use GPT4V for SFW captions, and that joy thing for NFSW captions?
>>101920040Try explicitly converting your values? idk
>>101920041same, shame that A1111 can't flux
>>101920088>sfw/nsfw status of finetuned model/lora output now relies on whether your prompting style pattern matches better to gpt4v or joycaption writing style and topic tendency
What the fuck is the point of the Guidance Scale? What was wrong with CFG? I'm a CFG guy. That's how I learned to use this shit and that's how I mean to keep using it.
>>101920040>>101920093 is essentially what i had to do but i also swapped the math node for int math from searge
>>101920106>What the fuck is the point of the Guidance Scale? What was wrong with CFG? I'm a CFG guy. That's how I learned to use this shit and that's how I mean to keep using it.that's what I do, I let GuidanceScale positive at 3.5 and I only change the CFGhttps://reddit.com/r/StableDiffusion/comments/1ekgiw6/heres_a_hack_to_make_flux_better_at_prompt/
>>101919431>not using bothNGMI
>>101920106FluxDev was trained to emulate CFG without actually having to do CFGThey used FluxPro to generate a ton of images at different CFGs then finetuned FluxDev with those images plus the CFG value used concated to the prompt so it learned what different CFG values look like
>>101920106Using real CFG makes it twice as slow
>>101919453With 6 GB vram I got OOM in Comfy but just werks in Forge
>>101920098>muh japanese retrowave look>but all the settings are a weird mix of californian and 50s americana
>>101920133*finetuned FluxPro into FluxDev
>>101920107>>101920093Yeah that's the issue. Both were int and the node was looking for floats.Thanks Anons.
>>101920088InternVL 26b is completely local and fits on a 3090, as it is better than 4V you don't need more.
>>101920133Sounds plausable. Source?
what in the sam hell is this shit and why is it taking so fucking long
>>101920160But it still begs the question, what changed? It was working perfectly this morning. >>101920107Did yours break today too?
>>101920152Forge is the new king. We should put him first in the general. Now all he has to do is to make his own node based ui to completely cuck Comfy.
>>101920172share your workflow
>>101920133Interesting. I guess I'll have to learn to use the Guidance Scale then.
>>101920172This is the same problem I have with forge. Even when I use SD 1.5.>>101920190>your workflowkek clueless
>>101920172Open your task manager and watch your RAM/SSD usage, maybe it offloads shit there
>>101920163The source is the description of FluxDev, the BFL code that runs it and my average IQ connecting the dots.
>>101920195I am clueless what are you doing?
Finally gave in and decided to install the Flux to see what it's like. It's 21 fucking GB. Fuck that shit. That's 10% of my hard drive. I guess this one is only for the rich kids who can afford the storage.
>>101920161If InternVL can't say that this character is eating with her foot, then it's not better than 4V, can it do that? https://files.catbox.moe/d0evak.png
>>101920180yes, and now workflows / seeds ive used before produce different outputs but that must be something on my end because i imagine others would have reported this. i used the math node for some resolution... math, but now even when i simply replace that with a primitive node, specifying exactly the res it would have been with the math, i get a different output. not sure whats going on.
>>101920228storage has evolved since 2005 bro
>>101920228>Finally gave in and decided to install the Flux to see what it's like. It's 21 fucking GB. Fuck that shit. That's 10% of my hard drive. I guess this one is only for the rich kids who can afford the storage.download the Q8_0 quant instead, it's twice as light (12gb) and the quality is similar to fp16 >>101918579
>>101920210NTA but in what instance could forge offload a 7 GB model to disk if I have 52 GB of RAM? Is there a setting to control this?
>>101920210ah yep, that's what's happeningi lowered the resolution a bit to where it would only use VRAM and it's zoomin' againguess i just need to find that sweet spot
>>101920156Sadly Flux cannot actually do vintage photography properly until someone makes a Lora. Not only it doesn't replicate the analog + film grain style, it also does not output period-specific hairstyles etc
>>101920152Something has been severely wrong with Comfy's implementation and I say this as a 24GB user it has never been smooth sailing. Always slows down when I start using my PC so I have to close everything and restart. Not even having completely cleared RAM fixes the issue apparently, so it seems to be a caching issue.
CCCOOOOMMMFFYYYYYY!!!!!!!!!!
>>101920250Thanks. Now that's more reasonable. 21 fucking GB. Seriously? Do we expect we run this in a server?
>>101920284it can be run on a 24gb card like the 3090/4090 but yeah... it's really reaching its absolute limit and I hate that, glad that quantization is a thing kek
>>101920284anon, it's not their fault you're a third worlder running mid 2000s hardware
>>101920270Also I have 64GB of RAM, I may upgrade soon but there shouldn't be a problem when I tend to have a ton available for Flux.
>>101920244Hmm, my outputs seem identical. It's only the math nodes that broke for me. But my workflow is pretty basic, just a gen and an upscale
>>10192029864GB of RAM may be enough for Forge, but Comfy isn't for poorfabs. Comfy was fast compared to Auto, yes, but only because Auto is a fucking mess. Comfy priority was always quality. If you want quality you go Comfy. If you want something fast and badly coded you go Forge. Simple as. I know in which side I am. I advice you to at least double that RAM if you want to experience what Comfy truly is capable of doing.
>>101920058taking a look at https://github.com/Z-L-D/Autocaption
>>101920298I got 32GB and don't have OOM issuesComfy does have some memory management issues, for example if you change the encoder it won't unload the previous one first and OOMs but that's something you rarely do
>>101920238Feel free to test it out https://huggingface.co/spaces/OpenGVLab/InternVL
hehehehe
>>101920325Wow imagine being this much of a faggot. You go comfy if you want nodes, you go forge if you want a webui interface, and you go a1111 if you are retarded or new.
>>101920268holy shit, someone already did it, I had no idea lolhttps://civitai.com/models/646686/japanese-photo-1980s-style-1980
nochekaiser881 started doing Flux loras, here comes the anime flood
>>101920389Can Joy do it?https://huggingface.co/spaces/fancyfeast/joy-caption-pre-alpha
>>101920363HF seems to be broken, usehttps://internvl.opengvlab.com/Their default is pro, their best model, change it to 26b in settings.
>>101920389>>101920394
>>101920363I can't ;'(
>>101920412>realistic uncircumcised state
>>101920324i reverted to a commit from a few days ago and im still getting different outputsfuck this gay ass earth im going to bed
>>101920389>>101920412This is not the foot prompt you described but NSFW...
>>101920389>>101920394this is joy, going to test this next: https://github.com/Z-L-D/AutocaptionThe image is a photograph featuring a penis with a cartoonish, exaggerated expression. The penis is circumcised, with a smooth, pinkish-red glans. The shaft is slightly darker and more textured. Above the penis, in blue, handwritten text reads "TEH REIL" in a playful, childlike script. The background is a plain white surface, likely a bed sheet or a tablecloth, which provides a stark contrast to the subject. The overall tone of the image is humorous and playful, with an emphasis on the absurdity of the situation. The image's lighting is bright, with no shadows, which highlights the textures and colors of the subject. The style is reminiscent of amateur, humorous digital art. The image is likely intended to be humorous or ironic, rather than serious or educational. The subject is explicit, and the image may not be suitable for all audiences. The image's composition and style are straightforward and unadorned, focusing attention on the central subject. The image's tone is lighthearted and playful, aiming to elicit a chuckle or a raised eyebrow. The image's explicit content and playful tone make it a memorable and attention-grabbing visual. The image's focus on the penis and the accompanying text creates a sense of whimsy and humor. The image's style and content are reminiscent of amateur, humorous digital art. The image's overall tone is lighthearted and playful, aiming to elicit a chuckle
>>101920363And... that's a failure!
>>101920444>a stark contrastGET OUT OF MY HEAD GET OUT OF MY HEAD GET OUT OF MY HEAD
>>101920444note I've modified joy caption to use the uncensored llama 3 and modified the prompt, but from testing it doesn't help too much
>>101920430to clarify, the math node is working again but im still getting different outputs and somehow they are more sloppa than before. it must be a custom node that changed but im not about to git revert every single one. goddamnit.
>>101920444Prompted this
>>101920468Certainly!
>>101920392God fucking dammit I was hoping he'd be filtered, now if I post my Loras to shitvit they'll sit beside his shitheaps
>>101920468It's important to remember that language is a tool we use to convey ideas, and sometimes common phrases can be a convenient way to express a thought. While "a stark contrast" might be a familiar expression, it serves its purpose in illustrating differences clearly. Let's not let the choice of words overshadow the message we're trying to communicate. The real value lies in the ideas and perspectives we're sharing, not just in how they're phrased.
>>101920412lmaoooooo, that's impressive
>>101920499If I didn't already have enough datasets to recaption to last me a lifetime I'd redo every lora he puts out but better, out of pure, unadulterated hatred
aren't we using flux1-dev-bnb-nf4-v2 ? that's 11GB, am I using the wrong model?
>>101920522that's the old thing, we're all about the new thing now
>>101920260i take it back, it's still fucking upgoddammit
>>101920522>aren't we using flux1-dev-bnb-nf4-v2 ? that's 11GB, am I using the wrong model?use Q4_0 instead, same size, better quality >>101918579
Come and get that nice late night/early morning bread...>>101920553>>101920553>>101920553
>>101920522we all hate nf4 and think it stinks! try city96's gguf quants instead. probably q4 if you're an 8-12GBer, q8 if you have 16+, though it depends on the speeds and resolutions you're wanting to work in. just start with the q4 and upgrade if you find the resolutions you want generate plenty fast enough for you. for comfy you'll also need his "bootleg" gguf loader for now. think they're natively loadable in forge.
>for some reason joycaption uses up exactly 4 GB vram per each of my GPUs and refuses to run on just one of them, no matter what I add to the code or modify in accelerate config I give up, use them both then
>>101920558what about the _1s ?
So does Comfyui have malware or not?>t. newfren
>>101920238>>101920460Are you sure you are using 4o or 4v?
im going to have an aneurysm
>>101920604the _1 are better than _0 abeit a bit bigger, it's something like this:Q4_0 < Q4_1 < Q5_0 < Q5_1 < Q8_0>>101920611>Are you sure you are using 4o or 4v?Yes I'm sure
