[a / b / c / d / e / f / g / gif / h / hr / k / m / o / p / r / s / t / u / v / vg / vm / vmg / vr / vrpg / vst / w / wg] [i / ic] [r9k / s4s / vip / qa] [cm / hm / lgbt / y] [3 / aco / adv / an / bant / biz / cgl / ck / co / diy / fa / fit / gd / hc / his / int / jp / lit / mlp / mu / n / news / out / po / pol / pw / qst / sci / soc / sp / tg / toy / trv / tv / vp / vt / wsg / wsr / x / xs] [Settings] [Search] [Mobile] [Home]
Board
Settings Mobile Home
/g/ - Technology


Thread archived.
You cannot reply anymore.


[Advertise on 4chan]


File: tmp.jpg (1.06 MB, 3264x3264)
1.06 MB
1.06 MB JPG
Discussion of free and open source text-to-image models

Previous /ldg/ bread : >>101828788

>Beginner UI
EasyDiffusion: https://easydiffusion.github.io
Fooocus: https://github.com/lllyasviel/fooocus
Metastable: https://metastable.studio

>Advanced UI
Automatic1111: https://github.com/automatic1111/stable-diffusion-webui
ComfyUI: https://github.com/comfyanonymous/ComfyUI
InvokeAI: https://github.com/invoke-ai/InvokeAI
SD.Next: https://github.com/vladmandic/automatic
SwarmUI: https://github.com/mcmonkeyprojects/SwarmUI

>Use a VAE if your images look washed out
https://rentry.org/sdvae

>Model Ranking
https://imgsys.org/rankings

>Models, LoRAs & training
https://civitai.com
https://huggingface.co
https://aitracker.art
https://github.com/Nerogar/OneTrainer
https://github.com/derrian-distro/LoRA_Easy_Training_Scripts

>Flux
https://huggingface.co/spaces/black-forest-labs/FLUX.1-schnell
https://comfyanonymous.github.io/ComfyUI_examples/flux

>Pixart Sigma & Hunyuan DIT
https://huggingface.co/spaces/PixArt-alpha/PixArt-Sigma
https://huggingface.co/spaces/Tencent-Hunyuan/HunyuanDiT
https://huggingface.co/comfyanonymous/hunyuan_dit_comfyui
Nodes: https://github.com/city96/ComfyUI_ExtraModels

>Kolors
https://gokaygokay-kolors.hf.space
Nodes: https://github.com/kijai/ComfyUI-KwaiKolorsWrapper

>AuraFlow
https://fal.ai/models/fal-ai/aura-flow
https://huggingface.co/fal/AuraFlows

>Index of guides and other tools
https://rentry.org/sdg-link
https://rentry.org/rentrysd

>GPU performance
https://vladmandic.github.io/sd-extension-system-info/pages/benchmark.html
https://docs.getgrist.com/3mjouqRSdkBY/sdperformance

>Try online without registration
txt2img: https://www.mage.space
img2img: https://huggingface.co/spaces/huggingface/diffuse-the-rest
sd3: https://huggingface.co/spaces/stabilityai/stable-diffusion-3-medium

>Related boards
>>>/h/hdg
>>>/e/edg
>>>/d/ddg
>>>/b/degen
>>>/vt/vtai
>>>/aco/sdg
>>>/trash/sdg
>>
Blessed thread of frenship
>>
I hope we'll be able to load stuff separately on Forge, so that I can choose the fp16 t5 with flux nf4
>>
We zooming now. Redoing all my Miku tests on Simple scheduler, apparently the fuckery on FP8 was because of Beta.
>>
NF4 is a speedup for VRAMlets and RAMlets only.
Which I am. Thanks llllllllllyasviel
>>
>>101830373
thanks for your hard work anon
>>
>>101830375
Yeah this seems to be true but I don't really get it
NF4 in Forge isn't faster on my 3090 than FP8 in Comfy but why isn't it? If I run a language model quant at 4 bits it will be faster than the same model in an 8 bit quant
>>
>>101830375
No it's faster for me on my 4080 too, but not by much. From 1.19s/it to 1.09s/it.
The main thing is it's not eating my entire system resource.
>>
I am no longer killing myself
>>
>>101830419
Sorry to hear that anon
>>
>>101830414
it's faster but not that much, and yeah I'm also surprised aswell, I guess that having different preicision for each weights makes the math more expensive than going for a homogeneous process like fp8, that's my 2 cents
>>
File: file.png (6 KB, 984x52)
6 KB
6 KB PNG
>>101830245
~10 sits for me on 2080, also 8gb
Now I'm curious how to get the nf4 quant out from a model. Been using a custom dev x schnell schizo merge from civitai in comfy yesterday and it's been working reasonably well at 15-16 steps, theoretically it should do even better in forge now
>>
>>101830414
Might be expanding to fp8 for computation, in which case we "just" need a custom kernel for it to be more efficient. No idea though.
>>
>>101830440
Meant for >>101830270
>>
>>101830414
it works well, text works well, I dont see any downsides to the nf4 model and only upside
>>
File: 1692300318743860.png (1009 KB, 1024x1024)
1009 KB
1009 KB PNG
>>101830455
example
>>
Is there a forge/A1111 equivalent to this?
https://github.com/asagi4/ComfyUI-Adaptive-Guidance
>>
>>101830455
text has always worked, fp8 or fp16
>>
>>101830476
yeah, but the quality is still good and less ram/vram is needed overall, so more people can run flux which is good
>>
>>101830373
Comparison from last thread while I gen some shit. Left is FP8, right is NF4.
https://imgsli.com/Mjg2NDIw
>>
>>101830447
>Body of 18 year old. Face of 8 year old
Typical AI shit
>>
>>101830320
I told it to draw an anatomically correct adult human woman as an ink drawing and it inserted eyes for her nipples AND belly button. Revolting. Multiple times. I fucking hate the sex hating freaks who build AI. Done with this piece of trash model.
>>
I know there's a comfyUi node that allows you to choose which kind of calculation you can get from the models, you could chose between the default, forge, A1111... I don't remember what was it
>>
>>101830539
search the manager
>>
>>101830511
Thank god Flux is censored for you freaks.
>>
>>101830468
Can't find anything. There's https://github.com/scraed/CharacteristicGuidanceWebUI but it's an alternative to Dynamic Thresholding though it's still interesting to test out
>>
>>101830519
>Body of 18 year old. Face of 8 year old
that's just what your average russian teen fitness influencer looks like
>>
>>101830534
Next time masturbate twice in a row before you play with AI.
>>
Need help again for forge, how can I change the model directory from forge? I can't see any option in the settings menu to do this. Is it even possible?
>>
>>101830553
It won't be soon and that's the terrifying part
>>
>>101830468
>>101830554
Actually there's https://github.com/guzuligo/CFG-Schedule-for-Automatic1111-SD
>>
File: 1693153028972472.png (1 MB, 1024x1024)
1 MB
1 MB PNG
yeah, nf4 is fine.
>>
>>101830580
I will laugh so fucking much when these AI companies will remove all children from their datasets, all midgets and short+petite people too. Then they can add NSFW back in and you will cry. No one will make "child finetune" for you.
>>
File: ComfyUI_01960_.png (1.08 MB, 1024x1024)
1.08 MB
1.08 MB PNG
So besides Miku and Goku what other characters does Flux recognize?
>>
>>101830614
based miku giving those poles what they deserve
>>
>Installing requirements.................
>>
>>101830580
You need to stop posting.
>>
File: ComfyUI_00464_.png (1.04 MB, 1024x1024)
1.04 MB
1.04 MB PNG
>>101830553
>>
File: superdave shocked 1.gif (2.44 MB, 400x224)
2.44 MB
2.44 MB GIF
>>101830650
brooo...
>>
>>101830589
I think it's less precise than Adaptive Guidance though, this one has a fixed value when adaptive just calculate some similarity to find the sweet spot

>>101830554
>an alternative to Dynamic Thresholding though it's still interesting to test out
interesting, I'll test it out, thanks anon
>>
>>101830580
SAI employee of the year?
>>
File: 1697453975558590.png (1.13 MB, 1024x1024)
1.13 MB
1.13 MB PNG
miku dressed as super mario in japan:
>>
retards still searching for some sort of sai conspiracy as a reason for pedo wiping this thread kekaroo
>>
>can't find Rust compiler
............
>>
>>101830650
OK this one is fucking funny though
>>
File: 1712457750961010.png (1.08 MB, 1024x1024)
1.08 MB
1.08 MB PNG
>>101830667
>>
>>101830686
So who's the pedo
>>
>>101830513
the problem is that we have no idea what it's supposed to look like, that's why a fp16 output would help there
>>
File: ComfyUI_01975_.png (1.15 MB, 1024x1024)
1.15 MB
1.15 MB PNG
>>101830667
How about Mario dressed up as Miku?
>>
>>101830698
Some mad troon
>>
>>101830562
Who wants to jerk off to mannequins lol. Flux has the exact same "sexy poses" for women that have been in the data set since 1.5 and they're all based on advertisements for literal sex dolls.
>>
>>101830718
I can't into fp16 so I couldn't tell you without putting my PC out of commission for 30 minutes.
You can have the prompt and try it yourself. Seed is 0
>A still frame from a movie starring Hatsune Miku, wearing a black crop top with the print "Sugoi!". She has a beautiful face with black goth smokey eyes eyeliner and mascara, black lipstick, large breasts. her navel is visible. she is wearing a thin o-ring choker and ripped jeans hotpants with a string thong underneath. her arms are covered in sleeve tattoos. She is lifting or pulling her top, the background is dynamic and vibrant.
>>
>>101830726
That describes many anon ITT
>>
>>101830723
>mamma mia
>>
File: 1716984048458958.png (1.11 MB, 1024x1024)
1.11 MB
1.11 MB PNG
>>101830723
miku + sanic hybrid:
>>
>>101830738
Sonic will always take over miku. Try to get a normal miku with sonic, she will always become a soniku
>>
File: 1692195480131710.png (1.11 MB, 1024x1024)
1.11 MB
1.11 MB PNG
>>101830746
this one is a bit more balanced
>>
>>101830732
>I can't into fp16
I'm not even sure there's such a fp16 that can be run into forge, like a safetensor file that has fp16 flux + fp16 VAE + fp16 clip_l + fp8 t5xxl
>>
>>101830738
credit were its due this one is kinda cute, like a nice OC.
>>
File: FD_00099_.png (849 KB, 1536x512)
849 KB
849 KB PNG
>>101830738
>>101830780
Mine from yesterday
>>
Oh comfy responded to the nf4 forge:
https://www.reddit.com/r/StableDiffusion/comments/1epcdov/comment/lhjxxg3/?utm_source=share&utm_medium=web2x&context=3
>I can add it but when I was testing quant stuff 4bit really killed quality that's why I never bothered with it.
>I have a lot of trouble believing the statement that NF4 outperforms fp8 and would love to see some side by side comparisons between 16bit and fp8 in ComfyUI vs nf4 on forge with the same (CPU) seed and sampling settings.
>>
>>101830790
>with the same (CPU)
why does the CPU matter for comparaisons?
>>
>>101830789
nice
>>
>>101830738
Prompt?
>>
>>101830790
classic comfy cope
>>
>>101830816
that one is simple like the mario one

Miku Hatsune dressed as Sonic the Hedgehog.
The setting is Tokyo, Japan.
>>
>>101830790
sour grapes
>>
>>101830825
Thanks. Interesting it automatically imparted that CG-esq style.
>>
>>101830790
>side by side comparisons between 16bit and fp8 in ComfyUI vs nf4 on forge
doesn't forge give different outputs to comfyui with the same settings?
>>
why is auto the most annoying faggot shit to install
>>
>>101830833
if you dont specify an artstyle it can be random, sometimes realistic, sometimes anime, sometimes pixar-like.
>>
>>101830809
>why does the CPU matter for comparaisons?
that's because in CPU calculation only, it's the same exact thing on ComfyUi and Forge, if we use the GPU there's gonna be some differences because each of them used different cuda and torch techniques for optimisation
>>
>>101830841
What are you talking about? I just cloned the git and ran it.
>>
>>101830790
Well, I'm not going to do FP16 because fuck that, takes way to long even on a 4080, but I am almost done with my FP8vNF4 comparisons.
>>
File: 1717655482190102.png (1.14 MB, 1024x1024)
1.14 MB
1.14 MB PNG
miku as princess peach:
>>
File: ComfyUI_02013_.png (1.04 MB, 1024x1024)
1.04 MB
1.04 MB PNG
>>
>>101830876
>Well, I'm not going to do FP16 because fuck that, takes way to long even on a 4080
ok I'll go for it, can someone give me a safetensor that has both fp8 t5 and fp16 flux inside of it?
>>
>>101830729
Some of these might be sourced from chink massage parlors idk.
>>
>>101830876
don't worry about it, the fact that NF4 wins on some of these comparisons or matches fp8 is already enough to justify implementing it. he is complaining there for other reasons.
>>
>>101830892
You can just run it from the unet with your weights set to fp8 for the fp8 bit.
>>
File: ComfyUI_00484_.png (1.4 MB, 1024x1024)
1.4 MB
1.4 MB PNG
Flux knows a bunch of mythological characters (Gilgamesh, Hercules, Ganesh, Shiva etc).
Also knows Wonder Woman, Batman and Robin, but not Raven or Starfire.
>>
File: 00155-655579979.png (1004 KB, 768x1024)
1004 KB
1004 KB PNG
This is apparently a cosplay photo
>>
>>101830790
>I can add it but when I was testing quant stuff 4bit really killed quality that's why I never bothered with it.
I can assume he was testing nf4 on unet models like SDXL, maybe on those ones it's overkill, but it seems like flux is more resiliant to quantizations (probably because of the DiT and because it's a really big model)
>>
>>101830902
Keep in mind the comparisons I did last thread are invalid which is why I am doing them again. I used the beta sampler and it seemed to cause some fuckery.
>>
File: file.png (1.27 MB, 1024x1024)
1.27 MB
1.27 MB PNG
>>
File: ComfyUI_00482_.png (1.53 MB, 1024x1024)
1.53 MB
1.53 MB PNG
I told Flux to generate Wonder Woman fighting Goku and it complied but made them both filipino for some reason.
>>
>>101830905
no I have to test everything on forge because comfy gives different pictures for the same seed and same models
>>
>>101830920
but it was used on both, it's clear now that these schedulers/samplers don't affect in the same way.
>>
>>101830887
does this shit support controlnet/img2img? Could literally just feed Flux the original meme image and replace the dude with miku.
>>
>>101830895
This is probably the only "sexy pose" that SD/flux "know" that is legitimately attractive but maybe I just have a thing for ponytails and lordosis.
>>
>>101830931
I love how the forge dev claimed nf4 is better than fp8 without proving anything, it's just "trust me bro" and they let us figure this shit out by ourselves
>>
>>101830975
img2img is inherent to all imggen models
>>
>>101830931
How about this one?
https://civitai.com/models/617609/flux1-dev?modelVersionId=690425
>>
>>101830939
Yes, but using simple gives a lot closer results, you will see when I finish
>>
>>101830997
no, it can't be 22gb, the fp16 flux alone is 23gb + fp8 t5xxl is 5 gb, it's probably a 28gb savetensor we're looking for
>>
>>101830981
what other "sexy poses" are there?
>>
>>101831012
https://huggingface.co/NeuroDonu/flux-dev-safetensors/tree/main
There's this one but it's a 33gb (probably flux fp16 (23) + t5xxl fp16 (10gb)), that's not what we're looking for
>>
>>101830981
Oh and these are okay
https://files.catbox.moe/ncrzxy.jpg
https://files.catbox.moe/e4fvxr.jpg
https://files.catbox.moe/xcoaak.jpg
https://files.catbox.moe/wm350w.jpg
>>
File: ComfyUI_02079_.png (1.62 MB, 1024x1024)
1.62 MB
1.62 MB PNG
Goddamn Flux really does feel like Dall-E at home.
>>
File: grid-0008.jpg (838 KB, 2688x3456)
838 KB
838 KB JPG
Miku comparisons are done.

MSI RTX 4080, i5 10400kf, 32GB DDR4 RAM.
Seeds: 0,1,2,3,4,5,6,7,8
Euler Beta. 20 steps. 1024x1024. Distilled CFG 3.5, CFG 1.

Simple Prompt: https://imgsli.com/Mjg2NDQy
Simple Text: https://imgsli.com/Mjg2NDQz
Complex Text: https://imgsli.com/Mjg2NDUy
Complex Prompt: https://imgsli.com/Mjg2NDUx
Complex Adherence: https://imgsli.com/Mjg2NDQ2
Style Adherence: https://imgsli.com/Mjg2NDUz

Seed 0 from each set:
https://files.catbox.moe/a58j5k.png
https://files.catbox.moe/mne0l6.png
https://files.catbox.moe/mogf5t.png
https://files.catbox.moe/5hoiy6.png
https://files.catbox.moe/25bqeq.png
https://files.catbox.moe/mtwdqs.png
https://files.catbox.moe/kgiirr.png
https://files.catbox.moe/zqm79f.png
https://files.catbox.moe/8dd748.png
https://files.catbox.moe/2hfg23.png
https://files.catbox.moe/tqoz6b.png
https://files.catbox.moe/qk15x1.png
>>
>>101831196
Fuck not Euler Beta, Euler Simple
>>
File: 1702271566355445.png (1.23 MB, 1024x1024)
1.23 MB
1.23 MB PNG
>Donald Trump punching a raptor
>>
>>101831209
Flux really struggles with violence. It's extremely hard to make it shoot, stab, punch or kick someone or something.
We need a violence LoRA
>>
File: 1702337923442116.png (1.38 MB, 1024x1024)
1.38 MB
1.38 MB PNG
>>101831218
it's still good though
>>
>>101831196
nf4 wins again imo, and even though there isn't as many broken limbs in simple, there are still some on the simple prompt for fp8, while nf4 is fine.
>>
>>101831232
https://old.reddit.com/r/StableDiffusion/comments/1ei5q6t/really_impressed_by_how_well_flux_handles_yoga/
>>
File: NotTheSame.png (3.65 MB, 3801x3235)
3.65 MB
3.65 MB PNG
Yeah... ComfyUi and Forge don't give the same output at all for the same settings, so we have no other choice but to go full comfy (if he implements it) or full forge to make some comparaisons
https://imgsli.com/Mjg2NDU0
>>
new to /ldg/ I mainly only use llm for textgen. I have installed comfyui and the basic example flow.

And found another workflow with a "realistic lora".

Anything else I should investigate or try for Flux Dev? I have a 3090
>>
>>101831244
You are literally using a different checkpoint in each one. Use the same checkpoint
>>
>>101831256
what would be the difference though? but yeah you have a point I'll try with the same checkpoint aswell
>>
>>101831019
There's like ... 10 of them total. And they've been the same ones since like 1.5, and I looked it up once when LAOIN was still online and like half were literally pulled from real doll ads and the other half were from Asian massage parlor ads. And they just haven't added any since. This is why the models based on booru tags properly just end up sweeping away the competition.

Now, with a model that is as good at comprehension as flux you can probably just prompt any type of pose you want but that also kind of depends on how much the model will just stop listening to you if it decides that what you want looks too much like sex or whatever.
>>
Okay I discovered the problem I was having with the errors it was spitting out.
Xformers is fucked.
If you're getting the "none type error" or other crap, just delete venv and let it rebuild itself.
Make sure you don't have the "--xformers" in your launcher or it'll shit the bed again.
Also update your Cuda, that seemed to prevent me from launching this.
I fucking hate software, pulled an all nighter trying to figure this out.
>>
File: NotTheSame.png (3.39 MB, 3801x3051)
3.39 MB
3.39 MB PNG
>>101831256
Nothing change, ComfyUi and Forge just can't be compared, they calculate shit diffrently
>>
File: 00180-3018208417.png (917 KB, 768x1024)
917 KB
917 KB PNG
Doesn't know Cirno very well.
>>
>>101831333
>ComfyUi and Forge just can't be compared, they calculate shit diffrently
it's still possible if you made a CPU only calculation, but who's gonna have the patience to do that lol
>>
>>101831232
>>101831221
>>101831213
>>101831200
>>101831178
>>101831364

How has this human filth not been banned yet?
>>
>all this shit about flux being fast
>it's slow as shit and uses barely 15% of my cpu
what the FUCK
>it's comfy's fault
I can't run forge, who do I have to rape to get this fixed
>>
File: 00182-3018208419.png (821 KB, 768x1024)
821 KB
821 KB PNG
>>101831278
If you pull fresh it doesn't even accept the --xformers argument anymore. When in doubt clone a new repository.
>>
File: Comfy FP8.jpg (747 KB, 1999x1999)
747 KB
747 KB JPG
>>101831239
It's only winning inside of Forge, because I am getting better results out of Comfy.
Same thing here, seed 0-8
Seed 0: https://files.catbox.moe/g8vgs1.png
Gonna run some more. I don't know how to save as a grid on comfy though so this might take longer
>>
>>101830790
Comfy may be too goodhearted to start a harassing campaign against Forge to force him to share his tech with us, but I'm not. Enough is enough. Forge is making us look like retards. If we can't get his tech by asking nicely, then we'll let him know what we think of him.
>>
>>101831382
>It's only winning inside of Forge, because I am getting better results out of Comfy.
what? you're talking about Comfy flux fp8 right?
>>
>>101831398
Yes.
>>
>>101831389
Nigga, you're kind of a creep.
>>
can we make foot fetish Dark Rey gens with Flux yet?
>>
>>101831389
hes really craving for attention, spamming this in b and here
>>
>>101831398
>>101831409
To clarify I am using the same model using the same prompts and the same seed inside of comfy and the comfy output on fp8 looks better than both the fp8 and nf4 in forge, at least from the one test I've done so far.
>>
>>101831409
yeah, comfyUi gives different results compared to Forge, and that's normal, they calculate shit in a different way, you should only stick to one software when making model comparaisons >>101831333
>>
>>101831410
>>101831422
How are you people STILL falling for this?
>>
>>101831428
are you using the fp 8 t5 or the fp 16 t5?
>>
>>101831450
sometimes is just him, when you don't interact with him enough. he just starts talking with himself.
>>
>>101831450
falling?
>>
>>101831450
What are we falling for exactly? The guy needs to be removed. This kind of shitpost isn't really ignorable.
>>
File: kid spammer.png (10 KB, 346x114)
10 KB
10 KB PNG
>the magician revealed his secrets on /b/
ahhh now i see how he's able to spam two fronts at once.
gotta kneel to this power level.
>>
>>101831481
Imagine falling for the oldest trick in the book. This are NOT real children, you gullible retard. He's using Flux to make them and trigger people like you.
>>
>>101831497
are you fucking dense?
>>
>>101831457
I am using the model fp8 comfy from comfy that he merged on day 1 of flux. It's the same one I used in Forge
>>
>>101831497
>Hey everyone! Look! I'm using this free tool produce one of the most socially taboo things in existence!
>I sure hope nobody sees this and collects it as a body of evidence! That would ruin EVERYONE'S fun :):):)
>But I just can't stop generating images of kids in their swimwear and posting them!!
>>
>>101831482
Fucking faggot will fuck this shit up for poorbros :(
>>
File: ComfyUI_00515_.png (1.7 MB, 1024x1024)
1.7 MB
1.7 MB PNG
>>
>>101831532
>using this free tool produce one of the most socially taboo things in existence
Epstein list now
>>
Is it better to run this locally from an ssd or hdd? does it put a ton of wear and tear on the ssd? is the speed of an ssd required?
>>
>>101831497
>This are NOT real children
high iq post
>>
>>101831471
oh shit, i guess i fell for it...
>>
>>101831549
>appealing to social taboo
>on an anonymous image board
...huh?
>>
>>101831512
No? They are not real kids. They are being made with AI. This is a thread for AI generations, einstein.
>>
>>101831538
You'l fuck it up for yourself bt not standing up for your rights.
>>
>>101831570
ugly gens are not allowed
>>
File: 1713462133437970.png (939 KB, 1024x1024)
939 KB
939 KB PNG
>>
>>101831577
What rights exactly are you implying here?
>>
I hadn't come regularly to this general in like a year and a half. Can't believe the little girls poster is still active.
>>
File: Capture.jpg (271 KB, 2384x1223)
271 KB
271 KB JPG
>>101830539
>I know there's a comfyUi node that allows you to choose which kind of calculation you can get from the models, you could chose between the default, forge, A1111... I don't remember what was it
It's that one anon, it will allow you to make ComfyUi calculate stuff exactly like A1111/forge
https://github.com/BlenderNeko/ComfyUI_ADV_CLIP_emb
>>
>>101831593
Do you think it's the same guy? ("I think btards call him teebs)
>>
>>101831599
What is the best system, bro?
>>
>>101827303

>>101827303
Sir i need proompt/word for that tummy
>>
>>101831604
the best system?
>>
>>101831599
Wow. So is this the reason why generations in Forge looks better than in Comfy? Thanks.
>>
>>
>>101831589
freeze peach
>>
flux is trash, I downloaded 25GB for nothing, fuck you all
>>
>>101831649
filtered
>>
File: 1720792203566499.png (819 KB, 1024x1024)
819 KB
819 KB PNG
Cought Imagen3 lackin' its good model but cencored like hell
>>
>>101831649
yeah but what's the alternative? it's the best we have so far
>>
>>101831649
Bro, you need to use a lora. You're missing out. There are so many cool loras out there for it now, bro.
>>
File: Capture.jpg (464 KB, 3840x1616)
464 KB
464 KB JPG
>>101831599
Doesn't seem to work on flux :'(
>>
>>101831640
Child pornography doesn't fall under free speech and realistic AI generations are equated to real thing.
Anyway, I think this guy is not an actual pedo, probably has some agenda.
>>
>>101831675
Listen man, could you at least post something good? This toddler shit aint it.
>>
>>101831683
His agenda is shit posting. But that aside, what child? How is it child pornography if it does not depict an actual child?
>>
>>
Are you guys fucking retarded, the prompts being used to test differences don't use weights.
>>
>>101831683
He's an annoying shitposter, but, again, these are not real children. Don't fall for it. This is what he wants.
>>
>>101831713
what do you mean?
>>
>>101831713
What do you mean? Like we should prompt Miku lifting weights? What difference would that make?
>>
>>101831723
You have to take your prompts to the gym to equalize their bench, squat, and deadlift before you can use them to compare models.
>>
File: 00012-4011938284.png (1.39 MB, 1152x896)
1.39 MB
1.39 MB PNG
nf4 20 steps 3070 52 seconds, 30 steps 78 seconds
>>
How fucking many proxies does this faggot have? I was struggling to find even one measly server that wasn't rangebanned (wasn't such an issue even a year ago)
No way he's using 4chan pass right?
>>
>>101831756
>no way he's a fed right?
>>
>>101831728
you shouldn't but because the prompts don't have weights nodes that make weight calculation the same between ComfyUI, A1111 and Forge are fucking useless.
This >>101831599 won't make the differences go away because we're not using weights to begin with.
>>
>>101831766
That would be hell of a retarded fed then, he ain't baiting anyone to join in
>>
>>101831779
>join in
join in what? its not illegal. the goal isnt to make you do it, it's to make the thread look bad.
>>
>>101831779
No but if you're trying to kill interest in something you behave exactly like what he's doing. He probably works for Microsoft or something.
>>
File: 00056-2015245386.png (792 KB, 1344x768)
792 KB
792 KB PNG
>>
>>101831800
based fed, gatekeeping for us.
>>
>>101831819
Thank you David Lynch
>>
>>101831786
>>101831800
he's posting on /b/ as well. maybe just terminally cumbrained neckbeard
>>
File: download (2).jpg (7 KB, 298x169)
7 KB
7 KB JPG
>>101831819
>>
File: flux_lora_1.jpg (355 KB, 1024x1024)
355 KB
355 KB JPG
It looks like flux can learn anime styles, at least when trying 512x512 training.
>>
>>101831875
flux can into girls with cocks when?
>>
>>101831875
this doesn't look like anything Flux can't already do
>>
>>101831875
qt kantoku girl
>>
>>101831196
Comfy vs Forge
SET 1: https://imgsli.com/Mjg2NDY4
SET 2: https://imgsli.com/Mjg2NDY5
SET 3: https://imgsli.com/Mjg2NDcz
SET 4: https://imgsli.com/Mjg2NDc1
SET 5: https://imgsli.com/Mjg2NDcx

MSI RTX 4080, i5 10400kf, 32GB DDR4 RAM.
Seeds: 0,1,2,3,4,5,6,7,8
Euler Simple. 20 steps. 896x1152. Flux Guidance 3.5, CFG 1.
FP8 Model hash: 275ef623d3
NF4 Model hash: c161224931

Performance for fp8 is identical in Comfy and Forge.

The above settings were what I used in Forge too, I fucked up my original post. You can confirm with the catbox's I posted so you know I am not shilling, just retarded.

Comfy Seed 0's
https://files.catbox.moe/g8vgs1.png
https://files.catbox.moe/l04z25.png
https://files.catbox.moe/gopn0t.png
https://files.catbox.moe/0caw66.png
https://files.catbox.moe/ppaax8.png

https://pastebin.com/GFNdv2vp
>>
>>101831875
god damn, thats pretty decent.
>>
>>101831705
Catbox?
>>
>>101831875
oh I know that drawing style
https://www.peakpx.com/en/hd-wallpaper-desktop-aasgt
>>
File: ComfyUI_00492_.png (1.19 MB, 1024x1024)
1.19 MB
1.19 MB PNG
>>
>>101831875
What hw did you use for training?
>>
>>101831903
What are the prompts tho? Would make comparing these better to know what was the original vision.
>>
File: 00065-891270002.png (1.02 MB, 1344x768)
1.02 MB
1.02 MB PNG
>>
> twice-divorced American woman with massive KNOCKERS
>>
>>101831903
I mean yeah, it is known that forge calculate things differently than Comfy, what are we supposed to learn here? >>101831333
>>101831599
>>
>>101831953
Catbox is there for the prompts.
Going to attempt an FP16 on some of them
>>
>>101831961
fffffffffuuuuuuuuuucccccccccck muh diiik
>>
File: 00067-3823626541.png (1.21 MB, 1344x768)
1.21 MB
1.21 MB PNG
>>
>>101831970
Seems to me like Comfy's implementation of flux is better than Forge, based on that.
Really need either a FP16 safetensor or for comfy to implement NF4 to really know if NF4 is better, or just a way for vramlets to use Flux
>>
>>101831961
Nearly average knockers.
Someone show Flux Milena Velba.
>>
File: Capture.jpg (86 KB, 1570x803)
86 KB
86 KB JPG
>>101831971
>Going to attempt an FP16 on some of them
is it possible to try fp16 flux + fp8 t5xxl on forge? I can't find such savetensor? can forge load models separately like on comfyUi?
>>
File: loss.png (38 KB, 601x317)
38 KB
38 KB PNG
>>101831888
You can't replicate subtle anime styles on it just by altering prompts.

>>101831905
I think it's still undertrained. Also, strangely it had a flat loss curve, but looking at validation images it learned a lot.

>>101831937
4090, 512x512 training, bs4, adamw_bf16. Now I'll go for the same run but with prodigy optimizer and after that I'll try 1024x1024 training on a rented A100.
>>
>>101831988
I want to know if NF4 is even capable of being used with controlnets and LoRAs.
>>
>>101832003
>I want to know if NF4 is even capable of being used with controlnets and LoRAs.
that's the right question, if it can't it's DoA
>>
>>101831997
I don't see a way to do that, no.
>>
>>101832017
fuck, now that I've spent some time on ComfyUi, I realise that A1111/Forge really lack some customisation, that sucks
>>
>>101832022
That's why they can both exist in the same space. WebUI is great for certain things, like XY's and inpainting, that shit sucks in Comfy.
I always use webui when testing my trained LoRAs, but for most other things I use Comfy.
>>
>>101831997
I know that you can load the VAE separately on A1111 but I don't think it's the case for clip, I could be wrong though
>>
Can someone explain what's the drama between forge and comfy, or if it's just anons here exaggerating everything to insane proportions?
Also what's stopping one from using other ideas? Both have public git, they're all built on open source stuff anyway.
>>
File: file.png (6 KB, 92x30)
6 KB
6 KB PNG
>>101831903
>>
>>101831990
> twice-divorced American woman with massive KNOCKERS, like MILENA VELBA in terms of the size of her boobs
>>
>>101832052
>Also what's stopping one from using other ideas? Both have public git, they're all built on open source stuff anyway.
Not for the nf4 thing, forge made it forbidden to copy that
>>
>>101832057
still lacking
>>
>>101832059
Watch me as I fork it
>>
>>101832059
That's not how copyright works.
>>
>>101832022
Forge lacks a lot of things compared to ComfyUi
>Only one unique positive prompt text box, no separation between clip_l and t5xxl
>Can't load models separately (VAE, clip, image model)
>Doesn't have a negative guidance
>Doesn't have an Adaptive Guidance extension
>>
>>101832052
OK so. Many moons ago in /sdg/ someone changed the bake to put comfy ui on top of A1111 and this caused schizos to have a melty because they hate comfy, then we got the thread split.
But it turned out this place was more comfy without all the avatar faggotry so a bunch of us migrated here.
Now any time Comfy does something that WebUI can't the same schizos have a melty about it. Then when WebUI (or a fork in this case) does something Comfy can't they cheer and shout about how they were right.
>>
>>101832102
Comfy's extensive and customizable features are worthless if they don't appeal to the vramlet hordes.
>>
>>101832110
why not both? a good Ui (forge) + a shit ton of features for the non casuals
>>
>>101832117
That's what stable swarm is. Simple UI with a comfy back end.
>>
>>101831532
>Hey everyone! Look! I'm using this free tool produce one of the most socially taboo things in existence!
>I sure hope nobody sees this and collects it as a body of evidence! That would ruin EVERYONE'S fun :):):)
>But I just can't stop generating images of kids in their swimwear and posting them!!

>In a board where people post Trump pictures because it is one of the only things the tool knows well
>>
>>101832109
>this caused schizos to have a melty because they hate comfy
hard to not hate that mf, he knew nf4 would make everyone's life better and he decided to not implement it anyway >>101830790
>>
>>101832109
>Now any time Comfy does something that WebUI can't the same schizos have a melty about it. Then when WebUI (or a fork in this case) does something Comfy can't they cheer and shout about how they were right.
OK so the drama is here, the devs themselves just do their own thing.
I don't get the idea since you can have both of them locally and it's good to have choice, but whatever, people gonna support the most inane shit I guess.
>>
>>101832127
you still need to touch the spaggheti shit before going into the Simple Swarm Ui though
>>
File: 48elkjcbuhhd1.jpg (42 KB, 540x346)
42 KB
42 KB JPG
>>101832052
Comfy and Ilyas were in a relationship for several years during university, so there's unfortunately quite a lot of animosity between them. Personally I hope they get back together, but the last time they met (during a meeting for Stability) there was a huge shitshow.

Comfy ended up leaving because Stability wanted to buy out web-ui and have the two frontends merged with Ilyas collaborating on the project on a contractual-basis. Comfy had various issues with this, the main one being that it would push out stableswarmui, who's dev is comfy's current partner. So comfy left, destroyed several stability servers on his way out (hence why the SD3 release felt so unfinished) and made ComfyOrg.

Stability are currently looking to poach Ilyas and still buy out webui, so many people speculate that Stability engineered this situation so they could get comfy out without breaking any contracts and then hire Ilyas fulltime to work on Stability UI.
>>
>>101832142
It pre-loads some spaghetti for you so you can just start proompting
>>
>>101832146
>It's all just fag drama
grim
>>
>>101830790
So... he's waiting for a proper comparaison before implementing nf4 is that it? How can we do that? We don't even have a fp16 flux + fp8 t5 to load on forge in the first place, it would be easier for him and everyone to implement it and then test the quality shit on his software instead
>>
>>101832146
>Comfy and Ilyas were in a relationship for several years during university, so there's unfortunately quite a lot of animosity between them.
https://www.youtube.com/watch?v=YaG5SAw1n0c
>>
>>101832146
Both are gay troons
>>
>>101832146
>destroyed several stability servers on his way out (hence why the SD3 release felt so unfinished)
I have a hard time believing this, sounds like an excuse.
>>
>>101832102
Don't forget the multi gpu feature that is only on ComfyUi
>>
A couple of FP16 comparisons
https://imgsli.com/Mjg2NDgw
https://imgsli.com/Mjg2NDg0
This shit takes forever for me and doesn't really prove anything so I don't think I will do any more until comfy implements nf4 or there's a full fp16 safetensor
>>
>>101832146
This is so many levels deep in stupid drama.
On a side note, it's very funny that the main best UIs for this are developed by gay people.
>>
>>101832146
Something something romantic relations at work bad idea something something
>>
>>101832190
>doesn't really prove anything
yeah, for the moment that's useless until we're able to load a nf4 model into comfy
>>
>3080 10GB VRAM
>32GB RAM

I can only get Flux to work when I increase paging file to 30+ GB. Memory usage spikes to 99 percent

If I buy physical RAM will it make any difference? I'm willing to upgrade to 64 or 128 but I suspect having only 10GB VRAM means I'll never be able to use Flux
>>
>>101832146
I guess at least it fuels their rivalry in working on their stuff.
>>
File: EXveWJlXsAAITJw.jpg (235 KB, 2638x1452)
235 KB
235 KB JPG
>>101832146
>So comfy left, destroyed several stability servers on his way out (hence why the SD3 release felt so unfinished) and made ComfyOrg.
such a kino fanfiction anon
>>
>>101832194
It's not surprising at all, iykyk.
>>
People really believed that comfy and illya were in a gay relationship?

Am I chatting with actual retards here? Or are you all larping with the fanfic poster?
>>
what do you goys think?
https://files.catbox.moe/1nek5p.jpeg
>>
File: FD_00003_.png (994 KB, 896x1152)
994 KB
994 KB PNG
>>101832194
>>101832226
They both probably have coding socks.
>>
>>101832000
Cool man let us know how it goes with A100
>>
>>101832256
proompt?
>>
File: 1722453528228895.jpg (132 KB, 768x1024)
132 KB
132 KB JPG
>>101832244
It's real anon. Most people don't know the true extent of their relationship prior to this. Why do you think their implementations are so similar and there's all this drama about code stealing?

It's because they worked on them together before parting ways and now they're each trying to claim parts of the original codebase as their own.
>>
>>101832180
It's just complete bullshit.
>>
File: FD_00006_.png (1.36 MB, 896x1152)
1.36 MB
1.36 MB PNG
>>101832256
>>
General for creeps. Yikes.
>>
>>101832342
general for cute waifu prooompters!
>>
>>101832363
Dafuck is a "waifu"?
>>
File: ComfyUI_Flux_6745.jpg (130 KB, 1152x864)
130 KB
130 KB JPG
>>
>>101832395
An anime girl that sweaty men form relationships with because they can't say no.
>>
>>101832395
Go
Back
Now
Leave
>>
>>101832398
he has a spooky face on his knee
>>
https://huggingface.co/StableDiffusionVN/Flux/tree/main/Checkpoint
I think I found it, the safetensor of flux fp16 + fp8 t5xxl (29 gb total), with that shit we'll be able to test the fp16 output on forge
>>
File: fluxgen2.png (1.01 MB, 1024x1024)
1.01 MB
1.01 MB PNG
my second ever flux gen. after it was done my RAM usage was still stuck 90 percent until i shut down comfyui. wtf anons
>>
>>101832395
Normies are not welcome here, you have to go back
>>
>>101832438
It keeps models in ram so that it can easily be reprompted.
>>
>>101832436
why not both fp16. it can run on a 24GB card with the --lowvram flag
>>
>>101832477
because we have to compare with nf4, and that one only has a fp8 t5, so with this safetensor the comparaison will have no biases
>>
forge says my device doesn't support torch/cuda but it should (4070) and i dl'd the newest cuda toolkit 12.6. the forge page says its for 12.4, do i need that specific version? i did run the update.bat and it said everything is up to date
>>
>>101832438
Are you trying to gen at fp16? (this is the default mode)
you can set it to fp8, which is what most of us do, and get good results still in a fraction of the time
>>
>>101832363
This looks like the bakemonogatari art style. Was this Flux or refined with something else afterwards?
>>
>>101832512
It's a Pony model, old gen.

https://civitai.com/models/54867?modelVersionId=539850

You specifically want the Amore(PDXL) variant, as the V2 removed the shaft look.
>>
File: ComfyUI_Flux_6771.jpg (141 KB, 1152x864)
141 KB
141 KB JPG
>>
File: ComfyUI_00989_.png (1.22 MB, 832x1216)
1.22 MB
1.22 MB PNG
added a bunch of plastic surgery terms to my prompt to really nail that "actress who hit the wall" look
>>
File: 1718129877761870.png (857 KB, 1024x1024)
857 KB
857 KB PNG
>>
https://github.com/comfyanonymous/ComfyUI_bitsandbytes_NF4

lol.
>>
>>101832743
holy fuck that was fast... let's fucking gooo now we can make proper comparaisons
>>
File: 3096724381.png (879 KB, 896x1152)
879 KB
879 KB PNG
>>
So all of this confirms bnb's superiority. What we 3090 Chads need is an 8bit implentation obviously.
>>
File: Capture.jpg (114 KB, 2238x408)
114 KB
114 KB JPG
>>101832743
To make this node work you have to do this on your cmd:
\ComfyUI_windows_portable\update>..\python_embeded\python.exe -s -m pip install bitsandbytes
>>
>>101832743
>>101832755
My body is ready. How the fuck do I use this thing?
>>
>>101832790
Thanks, installing dolphin porn now
>>
>>101832743
>#shamelessly taken from forge
>removes # Copyright Forge 2024
>applies AGPL
uh oh
>>
File: Capture.jpg (19 KB, 596x262)
19 KB
19 KB JPG
>>101832743
in case you can't find the node, it's this one
>>
>>101832817
You know what, Anon? I don't give a shit. So long as it works. Licencing is the bane of software development
>>101832790
OK installed the bits and bytes now what? Clone the repo to where?
>>
File: Capture.jpg (181 KB, 3159x766)
181 KB
181 KB JPG
>>101832868
>OK installed the bits and bytes now what? Clone the repo to where?
to ComfyUI\custom_nodes

I got an error though, the node doesn't work for me :(
>>
>>101832743
Wait wtf, I just realized it was Comfy who made this node, lmao he doesn't give a fuck about the licence or what? BASED
>>
File: 2024-08-11_00014_.png (2.25 MB, 1280x1024)
2.25 MB
2.25 MB PNG
>>101831196
>>
>>101832886
Update your comfy.
>>
>>101832886
update your comfyUI brotha
>>
>>101832743
STOP. This code is ILLEGAL. It belongs to forge.
>>
>>101832817
i think its because the forge repo is licensed under agpl so the license in the individual file probably doesnt mean shit
>>
>>101832886
Same
>>101832915
>ComfyUI is already up to date with the latest version.
>>
>>101832922
Not how it works. Individual files can have a different license.
>>
>>
>>101832918
>>101832915
Oh yeah right, forgot about that, thanks
>>
>>101832934
There's AGPL code from A1111 and GPL code from comfy in there so he can't put any code that is incompatible with either license in his code.

>>101832926
It might only work with that nf4 checkpoint file.
>>
>>101832934
Agpl is a special case.
>>
>>101832934
so you can just sneak a no commercial use file in an mit repo and sue them ok
>>
File: 3024127263.png (1016 KB, 1344x768)
1016 KB
1016 KB PNG
>>
File: 1723298979821335.jpg (142 KB, 3035x571)
142 KB
142 KB JPG
>>101832817
>>
So can we finally compare between this, fp8 and fp16?
>>
>>101832973
>It might only work with that nf4 checkpoint file.
I have it
>>
>comfy steals illyas code and removes credit in spiteful, petty troon fashion
>he fucks it up and it doesn't work
illyas just can't stop winning lmao
>>
>>101833008
Is it actually the same code?
Not that it matters much, someone would have done it and it's not a big deal.
>>
File: file.png (81 KB, 995x869)
81 KB
81 KB PNG
>>101833023
comfy didn't even change the method names lol, they're still referring to Forge
>>
I have lurked for some time and use kohya_ss and forge UI, I am aware of civitai and use their loras extensively and have grasped the basic limitations and good/bad practices in prompting and preparing training data. What I want now is to learn to create a good lora myself. I have tried multiple times with multiple approaches and the results are bad. 1st of all I tried to create a lora of a girl next door. I had a dataset of around 20 pictures of her, mostly high quality non-nude studio pictures with good lightning (she was posing for a photo set). Also some photos and selfies in bikini and even one topless shot, various poses. So the dataset was not perfect, as the lighting conditions and backgrounds were not varied enough, but it was a good start to get face and basic body shape correctly. After making a lora I started generating and only like 1/100 photos had face resemblance, the bodt shape was good though. I have managed to collect 3 good generated fully nude pictures with good face resemblance and added them to initial dataset. I also deleted some initial pictures which were obviously causing some artifacts (found which ones with extensive prompting). So the next lora with the new data had success rate of like 1/20. I have generated 5 more great pictures and added them to dataset to train again. The final lora has success rate of like 1/3 which is good enough for me. The problem is, I cant generate sex pictures with this lora. The model learned heavily the limited number of poses from the dataset, and as they did not include sex scenes, I cant enforce them now. Low strength means sacrifising the learnd face and going back to 1/100 success rate. What is my next step to overcome that? Regularization images? I have seen some datasets from civitai for characters, some of them are really limited compared to mine, and the prompting is really simplistic (i try to follow that), poses are never desribed, yet they maintain the flexibility somehow. What am I missing?
>>
>>101833038
>>101833008
It says the code is taken from forge in the readme lol nothing is being stolen.
>>
>>101833023
this is typical comfy retardation, he has some weird vendetta against illya that only exists in his mentally ill brain
>>
>>101833057
>What am I missing?
Your enter key
>>
>>101833059
Never said it is being stolen, the person just asked if it's the same code.
>>
File: ItsOver.jpg (1.05 MB, 3698x1496)
1.05 MB
1.05 MB JPG
It's ova... fp8 is closer to fp16 than nf4

>nf4
https://files.catbox.moe/p91tyu.png
>fp8
https://files.catbox.moe/thvzij.png
>fp16
https://files.catbox.moe/nmcpcu.png
>>
>>101833147
What about writing?
>>
File: file.png (3.32 MB, 1216x1662)
3.32 MB
3.32 MB PNG
>>101830658
>it's less precise than Adaptive Guidance
Well, it doesn't even work anyways. No matter what I tried, it produces exactly the same image at exactly the same speed. Maybe it's a forge compatibility thing, maybe it's a flux thing, or maybe both, but But Adaptive Thresholding itself does work as expected. Top - CFG1, bottom - CFG5 + AT, same settings as in Comfy.
I explicitly wrote "a real photo of" in the prompt btw, and CFG1 completely ignored that.
>>
>>101833147
OK but how did you get it to work?
>>101833147
https://imgsli.com/Mjg2NTA4
The output seems fine to me. Yeah it's worse but I would expect it to be.
>>
>>101833147
>>101833215

It can still do writing. My own comparison here:

https://imgsli.com/Mjg2NTA5

In other news, this "Flux.1-Dev Compact", which has a reduced filesize of 30% and is allegedly more accessible on less powerful machines, is identical to my FP16 gen.
>>
>want to try flux on forge
>downloaded the full package, run it, it werks
>time for update.bat
>now it spills spaghetti
tfw been downloading torch cuda with different versions 4 times
>>
>>101833247
Sorry forgot the link to the model I was talking about, here:
https://civitai.com/models/637170?modelVersionId=713385
>>
Next thread, if cloudflare lets me through...

>>101833248
>>101833248
>>101833248
>>
>>101833081
>OK but how did you get it to work?
https://github.com/comfyanonymous/ComfyUI_bitsandbytes_NF4

1) go on ComfyUI\custom_nodes and do this:
git clone https://github.com/comfyanonymous/ComfyUI_bitsandbytes_NF4

2) go on ComfyUI_windows_portable\update and do this: ..\python_embeded\python.exe -s -m pip install bitsandbytes

3) Update ComfyUi
>>
>>101833287
>https://imgsli.com/Mjg2NTA4
Yeah did all that, same error
>>
>>101832817
>as a custom node instead of the main repo



[Advertise on 4chan]

Delete Post: [File Only] Style:
[Disable Mobile View / Use Desktop Site]

[Enable Mobile View / Use Mobile Site]

All trademarks and copyrights on this page are owned by their respective parties. Images uploaded are the responsibility of the Poster. Comments are owned by the Poster.