Can we have shnell version 2?

@raidmachine132017712 I saw someone request it and it may come soon and sure it will be here too

Swarm should have asked you if you wanted to install the BNB/F4 code when you first tried to run the model. If it did not, try restarting and generating again to see if it pops up. If you did do the install and it's still generating errors, make sure you aren't using a LoRA and try again. Some LoRAs are erroring out in NF4 models. Most notably, anything specifically converted to run with ComfyUI will drop an error (whereas the original, unconverted versions seem to be running normally.)

I keep getting errors too. I updated SwarmUI and installed the additional items and I'm still seeing "All available backends failed to load the model." when I try to generate

nope...you are not. it consumes more vram/ram, equals slower performance on most of cards

on my 4070 the generation time is nearly identical, v2 may even be a tad faster

You're missing VRAM, I guess

I think we can inherit something like GGUF terminology.

You probably don't have enough VRAM. On my RTX 3060 (12 GB) the speed is about the same as v1. When you run out of VRAM it switches to RAM instead, which makes it alot slower. You should probably stick to v1 instead of v2, or find a way to free up VRAM (I think that between v1 and v2 it's only about 500 mb difference in VRAM).
Also I think that Forge might use a bit less VRAM compared to ComfyUI so you could try that.

I thought so too, but that was my mistake, on my 4070 12GB was even a few seconds faster with the same prompt and seed.

NF4 require "load nf4 checkpoint" node. It not work with regular load checkpoint.

Not yet...

Nope, you don't. The vae is baked into the model.

BNB stands for bitsandbytes, which is a type of quantization (like compression with some quality loss). NF4 stands for the size of the model. A "normal" model is released as a 16bit model, NF4 is mostly 4bit, with some of the most important data remaining in 16/32bit.

You can post the generated images to the corresponding checkpoint gallery area, or use ComfyUI-prompt-reader-node extension.

@jsineart bruh how did I miss that xd

You can just rename it yourself

You need Forge for it, right now. (https://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/981)

Nope, no auto1111 support yet

If you use nf4, 4070 will be faster since 12gb is enough to hold the model.

However, if you use fp8 it cannot fit in 12gb entirely, so 4070 will be much slower. (I use 4070 myself)

I cannot speak for whether 16gb is enough for fp8 though.

a higher side of vram helps in flux more,

@snap2887 Thank you!

Might be already started working on it........... :)

badly want flux for fooocus 😫

update already released guys . Forge Working great on every Flux Models + Loras and Upscaler.

@EKKIVOK That's true, it works in Forge without any problems. But Fooocus is so much sleeker when it comes to the overall application feel. I use Fooocus only because it's interface is so much less cluttered compared to the alternatives.

Upvote! Upvote!!

@RalFinger What app do you reccomend? I have A1111, Web UI Forge, Comfy UI and Fooocus

@RalFinger Thank you for the info. I'll try it in a few.

@RalFinger It works... but wow. So slow! Time taken: 1 min. 42.5 sec. for one photo. :(

@jaydoubleu_ with Forge now it only loads the model at the initial generation. However subsequent images will come out faster once it is loaded.

You need the Model and a Stable Diffusor Plattform like Forge that support Flux Diffusion, its a whole other category.

Yeah the Cleft Chin strong jaw line does seem to be a bit of an issue with the main FLUX models. Hopefully as the community finetunes it and adds in more image sets tit will be ironed out, however I'm kind of surprised that Black Forest Labs weren't aware of the bias it before releasing the model.

@Mirabilis Yeah, it's a little bit strange. I thought model should be truly unbiased in any aspect, no matter what is this aspect - face shape, breast size or anything else.
Well, let's see how it goes. Anyway, the progress is inevitable.

Also the faces always seem to have some weird coloring in the eye corners, or is that just me?

@JayNL You're not alone, I haven't noticed that until you mentioned, inner eye corners really have the same weird pale color in every female face. Looks like the image set the flux was trained on is flawed.

Came here to comment this. I've tried prompting different facial structures (e.g.: thin lips, round smooth chin, etc), given them names (e.g.: Anna, Jessica, etc. also actresses names), age (18yo, mid 30s, etc), ethnicities (Scandinavian, Latina), long/detailed vs short/keyworded prompts, but they all result in women having the same bone structure. It seems men also suffer from the "samefaceitis syndrome".
Perhaps it's a side effect of quantization (tested with Dev FN4) rather than a strong bias? It's quite inconvenient to load another model just to workaround it with inpainting/detailer.

@Nexdoor looks like some 'square-jaw-lover' forgot to add other female face types to a dataset. I doubt it's a side effect of quantization.

@muggzzzzz try to use some loras of celebrities with low weight eg 0.5 ;)

@sevenof9247 'The FLUX is the best model ever, but you should use Loras to get what you really want'. Does it sound funny? I'm pretty sure it does LOL
I think Flux is over-hyped, at least for now.

@muggzzzzz Which lora is better to use on women's faces? Nothing is gained. Women in flux are equally ugly, the names of the actors, styles do not apply to them.

@kiryanton930 I stopped using Flux because of all of the issues you mentioned, and went back to Pony based models.

i get all of reactions, thank you!

Only this fellow who pressed thubs down. With exactly what you dont agree? with random observation expressed as is?

It doesn't seem to be working. I tried using different LoRA files for Flux, but they don't seem to be compatible. There might be something that needs tweaking.

@Kiratinix current loras are meant for the original flux1.

@bananaisgod Flux D or Flux S? or original Flux base model

@Kiratinix I've tried different checkpoints seems it worked really well try mklanfluxdev Mklan-Flux-Dev-V1-FP8 CLip Vae Included - Mklan-FluxV1 | Stable Diffusion Checkpoint | Civitai

@pazuzu666 I tried it with Flux D, and it works with the LoRAs, but the strange thing is that some LoRA files work and others don't. Maybe it's something to do with the sampler method.

@pazuzu666 Well, this is not an NF4 model anymore. this means that it will work much slower on weak video cards

@Kiratinix what lora you were trying to use?

Noticed that when installing ComfyUI_bitsandbytes_NF4 [EXPERIMENTAL] via manager, there are alert that NF4 doesnt support LoRa

the LORA doesn't support this checkpoint is more likely.

Lora works in this model nicely. perfect result.

@AIStudio80 the don't work for me I'm using forge

i get a steady 1 second maybe 2 do to internet lag near flawless. with flux.1-schnell only 4 steps

@jconn2602399 in Comfyui or Forge?

Seems that flux1-dev-bnb-nf4-v2 which i have in Forge through Pinokio, it doesnt work with Flux LoRas yet, i tried with chloe grace moretz lora and it just didnt generated anything and got stuck at nothing.

@magmirant0011 it means that something is not being updated. the person who is currently engaged in FORGE updates various things there several times a day

--disable-cuda-malloc (not sure syntax for turbo, but guessing it is the same problem) is the only way I'm working with Lora's in ComfyUI.

Who is she? Link?

in webforge-ui

set GPU Weights (MB) to "7000"

@sevenof9247 i have it at default at 23000, is it wrong ?

So GUFF is better than NF4? can you tell me a GUFF model to try :)

@Frankcis RalFinger posted almost all of them, from Q2 to Q8

NF4 is still the fastest quant and gives comparable results to Q4_0. But Q4_K_S is already noticibly better

@Nelathan yeah they're not my words, I wanted to use NF4 again so I had to download bits and bytes and on the GitHub it said so, so I thought well then let's just go for GGUF, link here
https://github.com/comfyanonymous/ComfyUI_bitsandbytes_NF4

Try describing the background before the foreground/subject. If that fails, add more useless details like "the sky is blue". No idea why that works, but it does. The same flaw exists in every version, afaik.

@randomlettershdhtyxod 

yes i wrote an article ;)

https://civitai.com/articles/6892/anti-bokeh-prompts

That's 2.9s/it not 2.9it/s. It's completely different,and one is way faster than the other, 28s/it does not equal 28it/s on my gtx 1060, because that would mean i would be able to generate 28 steps in 1 second, and not take 28 seconds to generate 1 step.

@Gemini__ one step 2.9 sec

i feel like we have a similar setup but i havent tried 1 step. 30 steps take a little over 2 mins per image for me

you share no images no prompts so no one can help you ...

Use Euler Simple 20 steps

You did not follow Kanto yet?

DEV 2.90 it/sec ForgeUI 8gb RTX 4060Ti/32 gb ram image --> https://ibb.co/7SLsnvm

@sevenof9247 It's customary to provide information about your checkpoint and how to generate on it in the description for the release, rather than the other way around. Without doing this, one could be considered lazy or showing a lack of care in the things they release.

Again, the question is "What are the generation settings?", and replying with "you share no images" is not a reply to that question.

how long does it take to generate an image for you and what size?

@olvad 24sec/it, around 5 min per image at 896x1152. I come from 3D rendering so this is light speed.

nf4 does not support Lora

Under the Forge settings select 'bnb-nf4 (fp16 LoRA)' from the diffusion in low bit menu at the top and your Loras should run OK

Fairly sure that NF4 from Illyasavel has the T5, Clip_L and AE already included, hence the larger size compared to, say, Q4.

See here: https://huggingface.co/lllyasviel/flux1-dev-bnb-nf4 :

Main model in bnb-nf4 (v1 with chunk 64 norm in nf4, v2 with chunk 64 norm in float32)

T5xxl in fp8e4m3fn

CLIP-L in fp16

VAE in bf16

You can try comfyUI,Some picture uploaded here has nodes and prompt info,just downdoad one and drag it to the comfyui web page,you can get all you need,install the missing nodes by click the buttons on the comfyui mangager and check the model files that configed on the workflow, then you can make it work.

upload them to the model you used and they will be tagged with it

@dc802 Ah that's actually brilliant, I'll give that a try thanks!

@Unicore I didn't mention, but I already tried comfy and it wouldn't show either x) thanks though!

@Jhaku looks like it worked

@dc802 It has, thanks

Flux was trained on 12 Billion data points. SDXL was 3 Billion if memory serves. It's a larger data model so it requires more resources and is naturally slower on the same hardware, but as far as results, SDXL can not come close to the stuff I generate. I've all but left SDXL behind at this point. I run Forge SD on a 3090 Ti and it generates 3-4 images in about 1m 20s.

@cainezen Thanks to compare. On my system with a 3090 Ti 8GB it takes 5 min per image with this model. (24 cores P7, 128GB RAM)

If I use the regular full model, which is much larger, it takes 1.5 min per picture, or 8 s/it.

On this model the pictures are really dark, unclear and unsharp, means nothig is really recognizable (5 min to render a black picture ...). On the regular full model, everything is really clear and crispy.

I prefer ComfyUI to render, because all the FLUX outputs are black on my Forge installation. The speed is the same in ComfyUI but with good pictures on the regular models (FP8, Schnell and Full).

I found my problems in Forge: The designation of the sampler in the generated images is often "DPM++ 2M SDE", which is wrong. FLUX only works with "Euler", "Simple". Also the VAE "ae.safetensors" has to be specified.

The image data contains also a workflow, generated by Forge, which is mostly not working!

@dschonich Euler BETA is better.

@dschonich That is a good point. It's very important to follow the configuration guidance for Flux on the Forge SD Github page. I set mine up that way from the beginning. Also, I have 64GB of system RAM so I use "Shared" as my main method instead of CPU and it cuts about 15 seconds off generation time.

@dschonich sorry for responding so late but thanks for this, solved my issue I was having with it