please create a guff from the e4m3fn version too
can you please create a guff of the e4m3fn version too ? :
https://huggingface.co/Comfy-Org/Qwen-Image-Edit_ComfyUI/tree/main/split_files/diffusion_models
It doesnt make sense to create ggufs from fp8 versions when bf16 or f32 is available, because gguf does the quantization better and so the error is smaller with the same file size. Since our quants are based on the original model the quality is superior compared to ones based on fp8.
don't the e4m3fn versions also render with only 4 steps instead of 50 ? that was the case for the old model
that would be because they merged the lightning 4 step lora, you can just use that (;
oh the 4 step is the lora, but the e4m3fn version is 20 steps ? or am i misunderstanding the workflow notes? :
| Model | Steps | CFG |
|---|---|---|
| Offical | 50 | 4.0 |
| fp8_e4m3fn | 20 | 2.5 |
| fp8_e4m3fn + 4steps LoRA | 4 | 1.0 |
if you dont use the 4 steps lora, you should use 20 steps