I had an idea—what if I extracted the Schnell model from the Dev model? Could this result in a Schnell LoRA that retains some unique properties from the Dev model?
Turns out, it does! When you apply this LoRA at the end of your LoRA chain, it enables you to produce Schnell-like outputs from a Flux.1 D model with 4 steps... and you can combine other LoRAs if you like... but with not so perfect results... need more experimentation with that.
With this LoRA, you can even specify the FluxGuidance, just as you would with the Dev model, but with the added speed of Schnell.
Experiment and enjoy!
Description
1st release
FAQ
Comments (18)
It's nice that it takes 4 steps but in the end it looks like Schnell and not Dev
This is exactly the idea, you can transform a Flux dev model using a LoRA into a Schnell model… but with much more control. You can lower the LoRA strength to have less of a Schnell model effect, you can set Flux Guidance value, something you can’t do with Schnell.
With 7 step it looks a little bit better...
@Hegrie try setting the Flux Guidance between 2.5 - 3.5
This looks really promising!Any idea why my output is pixelated?
I am using: flux1-dev-Q3_K_S.gguf in ComfyUI
Tried guidance 2.5-3.5 and lowering lora strength.
It looks pretty good with guidance 3.5 actually...
I wonder if it could be better...Maybe it's because we need separate versions of this lora for differently quantized Flux DEV versions? Or maybe separate NF4 GGUF versions?
I'm using t5-v1_1-xxl-encoder-Q5_K_M.gguf in the CLIPloader.
I'll give it a shot on the fp8 DEV if I get the chance.
Great idea!
you need to do any merging/extracting, etc in full precision. after that it doesn't matter.
@Gore_Man Ah I see, that makes sense!
thanks for sharing. maybe adding some settings reccomendations or what you found to be good values would be great too!
Guidance is in double blocks. You can safely leave those layers alone (or move to schnell from dev). Would be interesting to see which layers have to migrate for the 4 steps to work on any dev.
Blanket subbing/merging leads to side effects.
This is actually made by extracting the difference between Dev and Schnell as a LoRA... not the same thing as merging models... At the moment there is no solution to extract a subset of the blocks.
@ainow I extracted blocks into another model fine. Saving it as a lora would just require some code modification. Finding what blocks are required is the first step. Maybe I should figure out how to make guidance into a lora. What tool/node did you use to save it?
@Gore_Man I used the the kohya_ss GUI code found in the sd3-flux.1 branch: bmaltais/kohya_ss at sd3-flux.1 (github.com)
@ainow Thanks. I will have to look into it because it doesn't expose the layers but it does go through them one by one.
Thank you. I have used the best lora of flux, which achieves a style, color and details similar to Dev, and sometimes even have better details than not using lora.
This is a great lora with 2 important features:
1. Gives flux dev the flexibility and more creativity that schnell has
2. It improves the characters lora adherence
i use it with 6 or 8 steps, and a lot of times it gives better result than the Turbo lora 8 step
This works with Flux Kontext too by the way.
I just tested it with Euler Beta and 4 steps, and the result is good. It's a bit blurry with the added elements. Is it possible to improve it? What parameters do you use? My machine is very old, and the few tests I'm running take me all day.
Would this work on GGUF models?
Details
Files
schnell_v1.0.safetensors
Mirrors
schnell_v1.0.safetensors
schnell_v1.0.safetensors
schnell_lora.safetensors
schnell_lora.safetensors
schnell_v1.0.safetensors
schnell_v1.0.safetensors
schnell_lora.safetensors
schnell_v1.0.safetensors
Hyper-Flux.1-Dev 4-step-Lora.safetensors
schnell_v1.0.safetensors
F.1-4-step-Lora.safetensors
Flux Dev_lora_Flux Dev 4-step.safetensors
Hyper-Flux.1-Dev4-step-Lora.safetensors
flux-dev-hyper-4steps.safetensors
Hyper-Flux.1-Dev-4-step-Lora.safetensors
flux-dev-hyper-4steps.safetensors
schnell_lora.safetensors
schnell_v1.0.safetensors
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.