These are merges of FLUX.1-Schnell and FLUX.1-Dev at different ratios/weights. The goal is a model which achieves near-Dev quality and ~Schnell speed. FP8 versions coming soon!
I recommend the following number of generation steps for these models:
| Model | Steps |
|-------|-------|
| 10:1 | 4-8(+) |
| 4:1 | 8-16(+) |
| 2.5:1 | 8-16+ |
New update 8/11: Comfy will now load these models! No changes to these files necessary. Make sure you update Comfy! See https://github.com/comfyanonymous/ComfyUI/commit/75b9b55b221fc95f7137a91e2349e45693e342b8
For more information on why these particular ratios, see my HuggingFace repo: https://huggingface.co/HaileyStorm/FLUX.1-Merges.
You may also generate images with the inference script there, which is based on the diffusers library. The nice thing is, it can use sequential CPU offloading, so you can speed up your gens regardless of how much VRAM you have (I mean, as long as you have... at least 5GB I think?). https://huggingface.co/HaileyStorm/FLUX.1-Merges/blob/main/infer.py
The licenses of both Dev and Schnell apply to these models.
Description
This is the Schnell 2.5 : 1 Dev ratio model.
FAQ
Comments (18)
any confirmation that 22gb work directly in comfy ?..
Comfy will now load these models! No changes necessary. See https://github.com/comfyanonymous/ComfyUI/commit/75b9b55b221fc95f7137a91e2349e45693e342b8
@HaileyStorm very nice thnaks ,
do you think it will run as same as flux merges i tried about the step speed ? i got some slowing down compared to sdxl , your model and any flux will run the same or this model here will improve that ? 'the step time how it takes'
@amazingbeauty the number of steps it takes to "finish" an image will vary when merges (the ratio). More Schnell means fewer steps to quality. E.g. I almost never go above 8 steps with the 10:1 model and for many things 4 is good.
@HaileyStorm yes thanks , but i still talking about '1step time it takes' not 'steps count' , if it's vary from flux merges , need something closer to regular sdxl 'the speed or single step' , idk it's flux or the merge i tried (not your model, didn't try it yet) , it's 'step speed' is very very slow like a stuck for something wrong..
@amazingbeauty it should run the same per step as Dev
exciting. i already found your models on huggingface, but they're split in 3 parts so idk how to handle them.
now i am looking for fp8 versions of these merges. coupled with t5xxl fp16 it would fit neatly in 24GB vram.
I am planning on doing fp8 merged, but: Comfy will now load these models! No changes necessary. See https://github.com/comfyanonymous/ComfyUI/commit/75b9b55b221fc95f7137a91e2349e45693e342b8
@HaileyStorm fp8 is dead. all hail nf4
EDIT: or maybe it isn't :D it's smaller at least
https://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/981
🤔
@tazztone nf4 is faster interference or just for lower memory ?
@amazingbeauty well i tested nf4 vs fp8 today in comfy. on 3090 both take 29sec (flux dev 20steps)
@tazztone hmmm, well, I'm taking my time about getting to the fp8 (sorry). Not sure about nf4...
@tazztone https://github.com/comfyanonymous/ComfyUI_bitsandbytes_NF4 Seems Guff rules now.
ye GGUF is the new NF4 :D
please help
Error occurred when executing CheckpointLoaderSimple: ERROR: Could not detect model type of: D:\ComfyUI_windows_portable\ComfyUI\models\checkpoints\flux1RatioMerges_101.safetensors
These models go in the unet directory i believe, rather than the checkpoints directory. You'll also need the text encoder and vae in the correct place.
I haven't run these in Comfy, but other users have.
@HaileyStorm thank you. i will try