Discussion of free and open source text-to-image modelsPrevious /ldg/ bread: >>101943936>Beginner UIEasyDiffusion: easydiffusion.github.ioFooocus: github.com/lllyasviel/fooocusMetastable: metastable.studio>Advanced UIAutomatic1111: github.com/automatic1111/stable-diffusion-webuiComfyUI: github.com/comfyanonymous/ComfyUIForge: github.com/lllyasviel/stable-diffusion-webui-forgeInvokeAI: github.com/invoke-ai/InvokeAISD.Next: github.com/vladmandic/automaticSwarmUI: github.com/mcmonkeyprojects/SwarmUI>Use a VAE if your images look washed outrentry.org/sdvae>Model Rankingimgsys.org/rankings>Models, LoRAs & trainingcivitai.comhuggingface.coaitracker.artgithub.com/Nerogar/OneTrainergithub.com/derrian-distro/LoRA_Easy_Training_Scripts>Fluxhuggingface.co/spaces/black-forest-labs/FLUX.1-schnellcomfyanonymous.github.io/ComfyUI_examples/flux>Pixart Sigma & Hunyuan DIThuggingface.co/spaces/PixArt-alpha/PixArt-Sigmahuggingface.co/spaces/Tencent-Hunyuan/HunyuanDiThuggingface.co/comfyanonymous/hunyuan_dit_comfyuiNodes: github.com/city96/ComfyUI_ExtraModels>Index of guides and other toolsrentry.org/sdg-linkrentry.org/rentrysd>GPU performancevladmandic.github.io/sd-extension-system-info/pages/benchmark.htmldocs.getgrist.com/3mjouqRSdkBY/sdperformance>Try online without registrationtxt2img: www.mage.spaceimg2img: huggingface.co/spaces/huggingface/diffuse-the-restsd3: huggingface.co/spaces/stabilityai/stable-diffusion-3-medium>Maintain thread qualityrentry.org/debo>Related boards>>>/g/sdg>>>/h/hdg>>>/e/edg>>>/d/ddg>>>/b/degen>>>/vt/vtai>>>/aco/sdg>>>/trash/sdg
>mfw
>>101948121
lmao thank you for using my horse meme image for the OP>>101948103fffuck, thanks anyway. Guess ill recreate it from scratch.
What were you thinking?
>>101948149debo no!
>>101948149advertiser-sama don't look
>>101948149
>>101948171Why did Joe spread a bunch of icing on her body and face
>>101948149>try the penis lora promptingwhy tho ani?
the fluxpro.art finally added seeds. Now I can test my gens agains pro properly.
>>101948184prompt?
>>101948220im sorry hombre that's a real image
>>101948093I noticed it the first time the other day doing Chika Komari gens. Noche's LORA would not ever get the ribbons right but Ibukimakisiko's LORA got them right surprisingly often.
>>101948213If you set safety to explicit, will it generate tiger balls?
When do you guys think they will release their text to video model?
We need new models like XL, sigma and black forest products every week.
>>101948261GOOD MORNING SIR
>>101948234baka art niggers always trying to pass of their real work as AI. put down the pencil and start prompting faggot
>>101948261Alright you start
>>101948235Im noticing the only good Stocking Anarchy lora on Civitai has a really hard time generating her iconic dress with the hair ribbon half the time, really considered just training it myself but that shit on 10 epochs will take almost 10 god damn hours..fug, maybe i can use my civitai good boy bucks but im not sure...
umm, bros??????
>>101948251It will generate whatever it is capable of generating. Unfortunately I don't think tiger balls are in the dataset.
>open new bread>see some chubby hairy dude's dickgays need to be removed
>>101948284>He pulled?
>>101948283kino
>>101948149'ick on 'eck
>>101948284What's your prompt?
>>101948311can we have a Bogdanoff Lora?
>>101948234
>>101948185dunno
>>101948335wish grantedhttps://civitai.com/models/150094/bogdanoff-twins
>>101948311>>101948332not a gen, just trying to train a loraseems to be something caption-related, more specifically related to special characters (0xe7 being "ç")do i REALLY have to manually clean all my thousands of auto-captioned pics of words like "façade" now?
>>101948284I'm having a shit load of unicode errors trying to tun that too. I just gave up because I can't be fucked.
>>101948373>do i REALLY have to manually clean all my thousands of auto-captioned pics of words like "façade" now?Write a script to convert to ASCII, not unicode or UTF8.
>>101948360need it for Flux tho
>>101948410train it on civitai then
>>101948376Decided to try it again and now it fucking works for no reason. I hate computers.
>>101948415It costs 2k buzz to train a flux lora on civit. So $2.
>>101948121He is /sdg/ jesus
>>101948662This really says a lot about society.Also, while I got training working, it seems to have frozen at picrel.Trying to run this on 16GB 4080. Any of you other "low vram" (can't believe I am a vramlet now) manage to get this going? I know some of you have trained LoRAs on less than 24gb vram. Would love some guidance.
>>101948698imagegen is fucking slow and i'm saying this as a 3090/24GB VRAM user
>>101948440
>>101948510Amazing
>>1019486984070 ti Super 16gb here had success training a few loras on Kohya using this config. https://github.com/bmaltais/kohya_ss/issues/2701#issuecomment-2294833735
>>101948662prompt for the ball gag?
>its only 600 tokens to generate a good lora for Pony on civitai>i have almost triple thathow the FUCK do people get away with making trash like nochenigger when its that inexpensive?Shit i need to triple check how the tagging works to make sure this goes good i might just upload the results then.Do Pony models absolutely need to follow booru tags, if i want to make totally sure it nails the exact outfit a character wears, should i make unique tokens to describe that outfit?
>$ sign in prompt>Comfy throws some EOF exception but continues with next prompt in queue anyway and hangs the systemAll of this 1girl making business is held together with wishful thinking, spaghetti and dried cum for glue
Flux or what it was got into the news because it's not censored and cucked enough. Is it still shit for anime style?
>>101948770>600 tokensmeant buzz. oops.
>>101948770Why do people hate on noche? His Loras are relatively decent for the average Lora.
>yfw 8gb vramlet
>2d,anime,1girl,office_lady,1boy,sitting_on_face,facesitting,from_side,assertive_female
>>101948763https://civitai.com/models/651337/ballgag-flux
>>101948753Oh with kohya? Did you use gui?
>>101948777checked. there is a lora but it's bad lol. I'd wait for a proper anime finetune.
>>101948829Yeah
>>101948833Thanks, I will give it a go. How long did it take per lora?
>>101948800makes undeserved $ from them, most of his loras are not great or just average compared to competing loras.in my case i get a special beef with him because his Stocking Anarchy one is pretty average to bad.
>>101948824it's the underscores, dummy
>>101948844>makes undeserved $ from themwho are you to say what is a deserved or undeserved amount? you're sounding very reddit right now.
>>101948862>criticism? must be from reddit!
>>101948840About 2 hours 40 minutes for 5000 steps
If any of you boys has a folder of the Bogdanoff twins, I'll throw it on Civitai and train a flux lora for them + post it. (after my stocking lora is done anyway :) )>>101948862>Um did you just criticize the lora maker that's very reddit chungushow about you make like a tree and get outta here, funny im making this response in the same post offering to make a free lora for a model i cant even run
>>101948875>>101948881thinking you know what value things SHOULD be is what makes you reddit brained, you retards
>>101948844>undeserved $So? People will make money off whatever they can. At least he's not selling generated images with tons of faults, he's not an e-whores who makes even more $ simply by flashing pussy to the camera either. If you truly cared, you'd duplicate his Loras but make them much better and sell them or release them for free to embarrass him.
ink sketch lora + basic miku prompt:
>>101948888>he holds no opinions about the world in fear that he might be wrong about themkek
>>101948888>>101948891why even cover him with these retarded arguments? Its perfectly normal to be critical of shit being sold as a product. What a waste of time.
>>101948905
>>101948914>>101948919you people are brain dead
>>101948931>he cowers in the corner shitting and pissing himself as he types the words "you people are brain dead" in concessionyou don't need to be like this.
>>101948939kill yourself, illiterate imbecile
>>101948939kek he struck your nerves
>>101948947>>101948952all me btw
>>101948817the extra minutes it takes to gen allows you to do other stuff in the meantime :^)
>>101948753Running this script I hit this error "network_module": network_module,UnboundLocalError: local variable 'network_module' referenced before assignment
>>101948919He makes good Loras, you try to stop him.
is there a node or extension in comfy that works similar to civitai manager? for loras and default instance prompts, I mean.
>>101949003Are you on the sd3-flux.1 branch of Kohya?
>>101949040Yes, I figured it out. The LoRA type was blank in the config. On to the next error
>>101949061Should auto pick it once you enter your model directories. Make sure you haven't loaded it in the Dreambooth tab... I've done that a few times...
can i run this on my chromebook
>>101949144yes
Does Flux strictly require Comfy?
>>101949159Works on forge
>>101949155where do i start nigga
>>101949159no
>>101949040>Are you on the sd3-flux.1 branch of Kohya?wait I'm not. Am I retarded or is there no branch of this? I can see sd3 scripts and that's all
>>101949168to clarify, he was joking.
>>101949195
>>101949186kys
>>101949224look buddy.. you don't wanna mess with melets just say im kind of a big deal around here
>>101949238delete this>>101949239teebs?
>>101949239Hello kind of a big deal around here, I'm Dad
>>101949184https://github.com/bmaltais/kohya_ss/tree/sd3-flux.1It's the bmaltais fork, I forgot this wasn't the main one, shit.
>>101949248Oh god teebs is here too?
>>101949255oh fug, thanks
>>101949206
im testing both guis, comfy is fine but sometimes in forge it will lag when unloading a model, how do you let it stay in memory, or unloading after each gen?
>>101949355>Crystalsong Forestcomfy
>>101949314I love it.>>101949355I love it.
Fuck, I pulled and now lora are not working anymore with gguf.AttributeError: 'GGMLTensor' object has no attribute 'tensor_shape'. Did you mean: 'tensor_split'?Please send help!
AttributeError: 'GGMLTensor' object has no attribute 'tensor_shape'. Did you mean: 'tensor_split'?
Miku but with a ghibli lora:
>>101949405Nice
>>101948110Do you think the success of Flux is a win or a loss for Stability AI? I've seen a lot of people saying that Flux killed Stability AI, but doesn't a rising tide lifts all boats? I think the future is bright for Stability AI. Flux will force them to release their better models.
>>101949195>>101949405>>101949435sick
>>101949405I love it.>>101949435I love it.Real bangers in this thread.This is the stuff of dreams my brothers. Imagine when these can all be animated with good quality.
>>101949456nah stability is toast.
>>101949475I like toast.
>>101949456>doesn't a rising tide lifts all boats? I think the future is bright for Stability AI. Flux will force them to release their better models.People have lost hope that Stability will actually release better models. They had to be aware that their former colleagues were about to release a competitor to SD3 and Stability hasn't seemed to do anything in response. The only way I see Stability surviving is if they can convince people to start training on their models rather than Flux
Is there a Flux guide (including requirements) for retards (me)?
Only way StabilityAI recovers is if they release SD3 8B, it beats Flux at what Flux does well and can do nudity out of the box.So it is impossible.
who's teebs
me
>>101949490>The only way I see Stability surviving is if they can convince people to start training on their models rather than FluxThey will. Flux (or Flush the toilet, as I call it) is too VRAM hungry and isn't as trainable as SD. Flux (or Flush the toilet, as I call it) is competition for Midjourney. The open source community will inevitably come back to SD once they realize it's better suited for our needs.
webcomic artist we coming for you next. Wait five, five months for the finetune and its a over.
>>101949537stop coping, Lykon
>>101949537nice bait
>>101949537nice meme
why is everything so hard to run these days
>>101949392For anyone else having this error, revert comfyui to 14af129c5509d10504113a1520c45b0ebcf81f14, latest commit broke GGUF lora.
>>101949537
>>101949578>latest commit broke GGUF lora.GOTTA LOVE UPDOOTING!
>>101949567Works on my machine
when will people stop making loras that override all faces in the image?
>>101949598stfu
>>101949608really i think the text kills the realism for these gens, It looks too perfect and obviously stands out from the rest of the image. It'd have to actually be affected by the room's lighting and actually look like marker.
>>101949613Kneel.
When generating an image, leave a piece of your heart in it. You are approaching perfection, approaching the divine.
>>101949613
>>101949608wow it really works!
>>101949627I'm making futa porn.
>>101949617some gens are better than others
>>101949643that's a whole lot better but still not quite there.honestly Flux 2.0 could probably do it out of the box perfectly, two more weeks boys.
another ghibli miku (lora)
>>101949674
>>101949674>>101949691That's a pretty big pussy.
Guess the website. Prompt is the blurb they use to advertise one of their video updates
pixar lora miku, got a cute pout face in this genhttps://civitai.com/models/650251
>>101949733
What's the biggest flux GGUF you can fit in a 12GB gpu?
>>101949775lmao
>>101949119This is great, catbox?
>>101949796for me Q4_1 is the best, Q5 seemed to sometimes fit and sometimes not fit, pushing the limit too much
>>101949809
>>101949811what's your speed with and without loras?
>>101949832~4.2 s/it on my 3060 without lora, seems to go up a tiny bit to 4.3 with6 s/it when i use bigger flux models that can't all fit in my vram
https://imgsys.org/rankingsflux got added to the rankings, and unsurprisingly flux-dev wins
>>101949869>Nude woman covered in magnetic tape from a VHS
Oh quants are about to get so much more confusing once I write the rest of the kernels.
>>101949894doing god's work anon
>>101949435fucking sick, how do you get that granulation effect?
>>101949894LET'S GOOty city
>>101949894city96 was always my favorite anon
>>101949894will there be a Q8_K?
>>101949894You fuckin' rock dude!
>>101949894nice, K quants are great. any chance for a K quant for the T5?
>>101949894oooooooo
>>101949912>illustration, (pointillism:1.3), (tight:1.1), ultra tight, super tight, very tight
>>101949931Nice colours
interesting, if I use q4/q8 models in forge, loras wont work properly, but they are fine in comfy. fp8 works in forge with loras, though.
>>101949937Probably not at first since that didn't come with a reference numpy kernel in gguf-py so I'll have to dig aroud the C++ logic to find out how it works, and there's more important tasks such as figuring out why the fuck the tensors still randomly fail when using LoRAs kek>>101949947Yes, comfy just added custom ops as an option for the text encoder + llama.cpp officially supports T5 so the quants will be standard too.
>>101949894is there going to be a ranking comparison once this gets released.Would be interesting which model produces the best outputs once the kernel gets out.
>>101949973see>>101949578
>>101949983no comfy is fine, I mean forge isn't working with q4/q8 when I use a lora, not sure why. I added the vae and encoders in the dropdown too.
>>101949982I'll probably make one eventually though I'm on a shitty 10GB 3080 and I'm pretty sure miku guy would kill me if I made him compare all of them kek.
>>101948753Is there supposed to be a flux script in here?
>>101949972tyty
>>101950029pls catbox i want to utilize this style
https://civitai.com/models/630820/flux-fusion-ds-nf4-fp4-fp8-fp16-4-steps-aio-and-unet-only?modelVersionId=705611It's gonna be such a mess, I don't think it's a good idea to categorise different quants as if it was another version of the finetune, maybe putting multiple quant options on the download button would be the thing to do instead
>>101949909She is so cute with BHA artstyle. Never thought i would want to see this mix.
>>101950050nice gen, catbox?
>>101950008I don't think so, I don't have one, I just loaded the config I didn't change any settings besides directing it to my folders and setting up the sample prompt. Make sure you're on the bmaltais fork >>101949255
so, why does comfy work with q4 + clip models + lora, but forge doesn't apply the lora for q4, it only works with fp8. is forge not working with quant models yet for lora? I get output but no lora.
>>101948110What is this horse from >>101948110
should i even bother with GGUF if i can run full bf16 in vram at 1.2s/it or does it offer a speed boost?i updated comfy to try GGUF after not doing so for a week and everything is broken (even after updating all nodes)
I have 6GB of VRAM and 16GB CPU RAM on a laptop, how do you run FLUX on forge?
>>101949894the model is 5gb, but it eats 6.8gb during inference, is that extra amount of vram consumption consistent?
I'm trying to run the gguf flux models for the first time, are these the two clip models I need?https://huggingface.co/openai/clip-vit-large-patch14/tree/mainhttps://huggingface.co/city96/t5-v1_1-xxl-encoder-bf16/tree/mainthe 2nd one is nearly 10gb!!
>>101950068Sure.https://files.catbox.moe/amb4iw.pngI found a new method using https://github.com/logtd/ComfyUI-SEGAttention for prompt adherence that's working out pretty well. It replaces the cfg entirely but you can still benefit from Adaptive Guider to disable SEG attention, like you can with the other method.
>>101950092I guess yeah, making the picture takes vram, especially at "high" resolutions like 1024x1024
>>101950075it's from this anon's gen >>101947888
>>101950105when I tried SEGAattention I thought it acted exactly like PrepNeg, it's 3x slower too right?
>>101950069Yeah definitely on this fork, ran the setup, pressed 1 to install gui, loaded script, set folders, pressed go, keep getting failures. This thing specifically
>>101949993I misread sry
>>101950055yeah I just got a bunch of civitai loras to test various styles, turned out pretty good
Someone save this for the next time someone asks how to use joy caption locally: (taken from /h/)You can clone joycaption from hugging face the same way you clone from GitHub (you can even use github's desktop app to do it if you feel so inclined)copy+paste your venv from forge or comfy or whatever you use into the cloned folder to save yourself having to build the venv and install all the shit it's requirements.txt is missingthen edit the app.py and change the>MODEL = line from hugging face to a local LLM of your choosinghere's a quant'd versionhttps://huggingface.co/unsloth/Meta-Llama-3.1-8B-bnb-4bit/tree/mainthough there are other uncensored ones that might be better suited when using for NSFW, it's worked well for me so faryou can write a .bat file to launch it if on windows (ask chatgpt how if you're not familiar)any errors just ask chatgpt how to resolve them
>>101950092Yeah it's like context with LLMs, the unpacked weights/actual image data/hidden states needs somewhere to go and they're in FP16.
>>101950146so what would be the equivalent to quanting the context cache? Do diffusion models have that?
>>101950050>>101950105>miku with black skin and dreadlocks skating
>>101950165Yeah that'd be it but no we don't have that. Pretty sure forcing those to be in fp8 would completely destroy the image quality. Some of the vram in my image is also probably from the vae being still loaded.
Is there a comparison between T5 fp8 and fp16?
>fp16why don't we have fp32? or is it just diminishing returns past that point?
>>101950123To be honest I don't know what Prepneg is. Gen times are about twice as slow now though. I like the results though. Tonemap seemed to lose some detailed compared to DynamicThresholding, but DynamicThresholding tended to end up with more artifacts like noisy images and stuff.
>>101950203there was in the very first days of flux, someone made a comparison with a photo of a woman and the fp8 T5 one made her skin more green for some reason
>>101949999>10GB 3080Lol, im using the same gpu. What are the odds. If you ever get to Q8_K if its ever possible that would probably be the most popular model for all us gpu lets.
>>101950221it's this: https://github.com/pamparamm/sd-perturbed-attentionbasically it has the same role as SEGAattention, it's supposed to give good results at CFG = 1
>>101950234>Mining coal with a battle axeGee no wonder why he is sad
>>101950231you can't fit Q8 or even Q5 on 10gb tho
>>101950140Really thankful you stumbled upon that mix. It so visually appealing to look at i can't put my finger to it. This just make Hatsune Miku 10x cuter
>>101950126No idea then
she is getting stronger!
>>101950221>Tonemap seemed to lose some detailed compared to DynamicThresholding, but DynamicThresholding tended to end up with more artifacts like noisy images and stuff.True, I tried for days to find the right parameters to recreate the magic unslopping effect that CFG + GuidanceNeg 10 + DynamicThresholding does, without much successhttps://reddit.com/r/StableDiffusion/comments/1euk2vw/some_xy_plot_i_made_to_try_to_understand_the_cfg/Maybe a method that will have the DT effects without the artifacts exists and we haven't found it yet, or it still doesn't exist and we are searshing in the void kek
>>101950264i managed to still use the models with forgeui. It just took me 4 to 9 minutes of wait time to gen a single image 1024 x 1024 last time i tried. Just not really worth the wait time when i could do the same with Q4 faster.
>>101948259They said it's coming soon, I imagine the next month or so.
>>101950249well shit, you're right. just tested and got almost identical gens. thanks for letting me know.>>101950293So that's still your workflow then Miku-anon? RIP. I'm probably gonna go back to it because these double gen times suck
>>101950307desu I doubt we'll be able to run it on our machines, unless we use ultra optimized quants like Q6_K then maybe?
>>101950252yeah flux seems to think pickaxe = axe sometimes kek>>101950307man that sounds like it will have horrific gen times, but still sounds cool
>>101950171>insert autistic screeching about how he can't gen it without a hack
>>101950080It can be useful to run the Q8 or Q4 version if you want to keep some VRAM free for a local LLM to help write prompts
>>101950171>>101950382Funny thing is, I'm not even that anon. Just the combination of conflicting features on Miku's base model makes a good way to benchmark prompt-adherence.
>>101949868>>101949796i'm not sure why, but the full size models work on mine without any offloading at all, with 12gb vram ~3.7 iterations/second
Just to think.. In a year's progress, we'll be running audio visual and text all on one machine at perfect accuracy....You're huffing the same hopium too right?
>>101950421Even my 4090 can't run weight_dtype: default, lmao.
>>101950421>without any offloading at allnah, comfy intelligently offloads when necessary and doesn't necessarily fill every last byte of VRAM first, i am also able to run fp16 on my 3060 like thatthough for me it's more like 6 s/it, not sure why yours is faster
>>101950249>>101950221My b I gave you the wrong one, it's PerpNegGuider and it's native on ComfyUi https://perp-neg.github.io/https://civitai.com/models/625042?modelVersionId=706228
This is a photograph capturing a man squatting on the ground in a bustling market street, with the iconic Taj Mahal in the background, shrouded in a light haze. The man, of South Asian descent, has a medium build and is dressed in traditional attire: a light beige kurta (long shirt) over a dark vest, with loose, beige trousers and flip-flops. He has short, dark hair and a neatly trimmed beard. His expression is relaxed, and he is smiling slightly.The street is narrow and crowded with makeshift stalls and vendors, covered with blue and brown tarpaulins. The ground is dusty, with scattered debris and small objects, including a few plastic bags. The stalls are low, with a few motorcycles parked between them, adding to the chaotic yet lively atmosphere.The Taj Mahal, a grand mausoleum with its distinctive domes and minarets, stands prominently in the background, its ivory-white marble glowing softly against the hazy sky. The surrounding buildings are a mix of modern and traditional architecture, with some high-rise structures visible in the distance, blending into the misty atmosphere. The overall scene is rich in cultural and historical context, capturing the essence of a busy market day in a historic city.
>>101950421I'm using q5.1 and I'm getting 5.2 s/t for 1024x1024 pics. What am I doing wrong.
>>101950439I have a 4070 Super and I can run the fp16 model and clip/T5 no problem with ComfyUI. I think it has something to do with ComfyUI's lowvram mode that allows it.
>>101950458im almost in tears at my stocking LORA from civitai not going very well on the 20th epoch, im gonna throw this guy's prompt into AutismMix confetti and see what it gives me.
>>101950461i have no idea, i'm extremely stupid and i just use the default settings from the noob guide http://comfyanonymous.github.io/ComfyUI_examples/flux
>>101950434yep, we will also gets img to vid locally aswell. Give it only a few years and people will be making fully feature films from their own home.No more hollywood or bollywood films polluting the medium with propaganda now that you can have them home baked.
Flux, that's not what I meant by dangling pair of legs...
>>101950458what did it mean by this?
>>101950464Hm I guess the implementations are different. Sick results though, will try and tune this and see what I prefer.
>>101950503what prompt did you use for the movie poster.
>>101950527Oops wrong post>>101950450
>>101950421>full fp16 model>12gb vram>~3.7 iterations/secondExcuse me? I hope you mean 3.7 s/it instead?
>>101950528>A poster for an animated 3D Disney movie. The poster says "DISNEY Pixar presents EPSTEIN" at the bottom, with the stylized Disney and Pixar logos. The rest of the poster shows an animated depiction of a dark jail cell. Behind the bars, a man's legs can be seen dangling from the ceiling, subtly illuminated. The legs are clad in gray pants and black shoes. There is an empty bed in the corner of the cell, and the floor and walls are made of concrete while the bars are cold hard iron. The legs hang in the upper half of the view, and the rest of the hanging body is out of view. The image implies someone is hanging from a noose after committing suicide, but all that can be seen is the dangling legs.
I dont get it, loras dont always work in forge with a quantized model but it's fine in comfy, fp8 always works in forge for loras. but I wanted q4/q8 cause I have 16GB.
>>101950322>>101950346If it's good, people will find a way to bring down requirements and speed up generation, just like with Flux. Maybe they'll also release a distilled model like Schnell, if that's possible for video gen. I'm pretty optimistic based on the past few weeks, but we'll see.
>>101949982Do you need it? You can make a direct beeline from how these prompts perform in LLMs to do a fairly accurate prediction how badly things degrade the smaller of a quant you use. Based on the current image, I am willing to say that Q4_K_M is going to be the one that probably will be the least compromises and the one everyone shoots for while Q6_K is probably going to be the threshold where it will still adhere to the prompt. I personally like Q5_K_M with LLMs though but it's possible there's going to be enough degradation of the diffusion model that people won't go for that quant.
>>101950557Give it a couple more days before people settle on the right way to quantize the loras. Certain implementations will be better than others
>>101950551sorry, yes, 3.7sec/it. in my defense, i did say i was retarded
>>101950572is this resampled through a pony model, or have flux LoRA's come this far this quickly?
>>101950552Thank you so much
>>101950527There's even PerpNegAdaptiveGuider if you want to combine PerpNeg with Adaptive Guidance (for the boost speed)https://github.com/asagi4/ComfyUI-Adaptive-Guidance
>>101950577fp8 works fine but occasionally there will be lag when unloading a model, I want a toggle to keep it in memorycomfy is good but I need a tool like civitAI helper for default instance prompts, as it isnt always obvious.
>>101950607Oh nice, thank you.
https://civitai.com/models/657252/fluxstanza?modelVersionId=735368a costanza lora, we are in the new age of memes
>>101950591just flux w/ a lora
>>101949894Any chance of t5 quantization?
>>101949894Well, _K quants implemented. Thank god the llama.cpp guys had reference numpy code for most things.Q6_K is for some reason faster and better on SD1.2 than Q8_0 lol, then again, it's SD1.2.The actual c++ code doing the quantization is questionable and it's hard to find analogies to the keys they use, trying to figure that out before uploading them. Will also have to quant from FP32 because I didn't add BF16 support lol.
>>101950670damn, very nice.
No CFG used, close to getting the benchmark that people keep using
>>101950572That's a cute style, is that the LoRA?
>>101950569i guess only time will tell. Once we get these models on hand we will know for certain which is the best gens for least gpu usage.
>>101950701nice, care to give your workflow?
>>101950701Try enforcing an uncharacteristic style like "50s comic book" or something, that's where I ran into issues
>>101950458>>101950651
>>101950624>>101950527what PrepNeg parameters did you use to get that insane result?
>>101950728fucking visionary over here
>>101950719https://files.catbox.moe/6djyrj.png>>101950727I'll try that out
>>101950728kek
>>101950692so it's normal for Q5 to be slower than Q4?
>>101949894>>101950692I kneel.
>>101950732Default PerpNegGuider, cfg 1.0 and neg_scale 1.98. Seems really unstable though since I didn't prompt for the butterflies, just got lucky on the first gen. I can give you a box if you want
>>101950748>Boomer promptingMany such cases with Flux :(
>>101950774I used an LLM to turn my non-boomer prompt into a boomer prompt. I hate that it works so well but I enjoy the end results.
>>101950774Is it really boomer prompting when it is just how it works? It's just describing the image.
>>101950774kek i am doing the same thing, write the prompt in tags in clip l, then ask gpt to turn my tags into a boomer prompt for the t5
Q2_K is, as expected, completely retarded.>>101950758Yes I think my shitty tensor block access code was slowing it down, I tried to improve that as well but who knows if it did anything lol.
>>101950794Well, at least it can still spell!(prompt that please)
>>101950794MsPaint Flux
>>101950787>It's just describing the image.Adding both "Black African" and "dark skin" is equivalent to (dark skin:2), basically you're just spamming that token to force Flux to consider it, that's not natural at all
>>101950794Q2_K = soul
>>101950769>Default PerpNegGuider, cfg 1.0 and neg_scale 1.98. Seems really unstable though since I didn't prompt for the butterflies, just got lucky on the first gen.yeah maybe neg_scale 1.98 is a bit too much, on the civitai example it's at 1.5, but yeah, seems promissing indeed
>>101950769>I can give you a box if you wantSure, I'm interested on that one
>>101950804>basically you're just spamming that token to force Flux to consider it, that's not natural at allIs it any less natural than hacking back in CFG and using a negative prompt to "force" the model not to include it?
>>101950799Wonder if I could whitelist some crap to make the quality degradation less horrible. A 3.8GB file outputting trash is less useful than a 4GB one that doesn't mangle the output.
>>101950835>me every day for the past week or so that i've been lurking these threads
>>101950832https://files.catbox.moe/qips4g.png
>>101950839unironically better than SD3 still
>>101950839honestly not too bad at all, still better than the equivalent SD lmao >>101950848
>>101950833I much prefer adding a CFG hack that works and then not touch it again than spending the rest of my life writing stuff like "A drawing of Miku as a black african, it means she has a black skin, black as is the absence or complete absorption of visible light... please flux consider that concept of blackness!!!"
>>101950847thanks anon
>>101950847too bad you don't have a seed node so I can't reproduce your results
>>101950926Seed is:438710575208934
>>101950960oh it was that red node on my side, kek, is there a reason you went for that custom node instead of the regular seed node on ComfyUi?
>>101950839Can you fix your node with recent comfyui updates?>https://github.com/comfyanonymous/ComfyUI/commit/bb222cebreak lora with gguf>https://github.com/comfyanonymous/ComfyUI/commit/4f7a3cbbreak loading model
>Here's your lora broOnsite training was a mistake
teal color power aura miku
>>101950976Does comfy have a standalone seed node that you can link to all your inputs and can randomize and fix on a button press? Honestly never bothered to look lol, rgthree was one of the first custom nodes I downloaded since he has a plethora of utility stuffhttps://github.com/rgthree/rgthree-comfy
>>101950982cant wait for the female brit lora. Man is making some progress.
>>101950982looks about right
>>101951004>Does comfy have a standalone seed node that you can link to all your inputs and can randomize and fix on a button press?no it's not that sophisticated kek, but desu if I want to change a seed I just do a +1 kek
>>101950982I block every civitai users that make retarded loras like that, made my life way easier as there aren't that many people making loras to begin with
>>101951059tbf im used to using the default KSampler nodes that just have a seed input
