Unet (without clips ans vae) bnb-nf4 version of the merge of Flux.1-schnell and Flux.1-Dev
Comfyui node for unet nf4:
https://github.com/DenkingOfficial/ComfyUI_UNet_bitsandbytes_NF4
I used the tip from comfy.org for merging just the Flux.1-Dev double_blocks (MM-DiT) onto Flux.1-Schnell, giving you a higher quality model that still runs in just 4 steps!
https://blog.comfy.org/august-2024-flux-support-new-frontend-for-loops-and-more/
All in one:
Clips(t5xxlfp16, L) and vae are included, you can put it in comfyui checkpoint folder and use comfyui default WF
Work with Forge 2.0.1
Description
Flux.1-Schnell merged with Flux.1-Dev
FAQ
Comments (41)
Hey thanks for this :)
You're welcome
would it be possible for you to provide an fp8 version of this merge? also to merge the two models locally, how should I load models in the load diffusion model node given they both are divided in multiple parts on huggingface. do I need to somehow merge all parts of dev and schnell into one safetensors file each and then load them into load diffusion model node for merging?
This is the fp8 version with clips and vae included you can use it with the default WF of comfyui, put it in the checkpoint folder not unet folder
In this news: https://blog.comfy.org/august-2024-flux-support-new-frontend-for-loops-and-more/ comfy.org provide the WF for merging
@jice which is better at 4 steps , your model or the merge that mentioned on comment from hugging face ? and whats wrong with it ?
@333helix which is better 4 step model ? the model you linked from hugging face or this model here on this page ?
@amazingbeauty did you find an answer? also looking for this
cool. i thought schnell was just a turbo version of the dev version. Maybe merging the turbo version and main version gives better quality models in general? :D
Yes, merging the Flux.1-Dev double_blocks (MM-DiT) onto Flux.1-Schnell, giving you a higher quality model that still runs in just 4 steps!
this is the unet all all in one file ?
Yes, clips and vae are included
@jice Okay dumb question, this also goes in the Unet folder correct?
@olternaut No, now it goes to regular checkpoints place , laod like any checkpoint
@olternaut It also works in the unet folder, but it is not necessary to load the clips and the vae separately.
Thank you! Finally able to run Flux with 32GB ram and 6GB Vram on simple Comfyui. 4 passes = 195 seconds. 512 x 512
Are you sure about the license being Apache 2 since it was merged with the Dev model??
You are right, i changed the licence
what vae do you use? i got error, All available backends failed to load the model.
The vae are included to the model, put the model to the checkpoint folder and use the regular comfyui load checkpoint node
Thank you! This is perfect!
thx for your feedback
Critical piece for getting this to work is:
1. Update ComfyUI. (You need the Flux Guidance options for the conditioning / prompts)
2. Put a Flux Guidance block between each prompt, negative and positive, where it connects to the sampler.
Set the CFG to 1.0 also.
That should work. It did for me, and this was the first Flux model I found that I actually could get to fully work.
Flux guidance not work with Schnell version, now i use Forge 2.0.1
can you make the nf4 version?
A gguf version has been released: GGUF: Flux.1-Schnell Merged with Flux.1-Dev - Q4.0_v1 | Stable Diffusion Checkpoint | Civitai
@elyzionz1 nf4 version is there
That's great. Well done.
Thx for your feedback
Doesn't work in default comfy WF. getting error:
Error occurred when executing CheckpointLoaderSimple: module 'torch' has no attribute 'float8_e4m3fn'
updated to latest version of torch, same error
***Solution was update comfy, but update+dependancies
Working good for me, try to update comfyui?
@jice Comfy's up-to-date, what kinda gpu do you have and what version of torch are you using?
@Scew 4060 ti 16gb, pytorch version: 2.2.2+cu121 for comfyui but I use forge 2.0.1, 99% of the time
@jice just ran the wrong update for comfy, needed the update+dependencies. Seems to be working now ^.^ #seemsToAlwaysFindAWayToMakeItHardOnMyself. lol
Details
Files
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.