FP8-scaled LoRA for Qwen-Image-Edit-2511

[menahem121]

Hi, is an FP8-scaled version of the Qwen-Image-Edit-2511-Lightning-4steps LoRA planned?

At the moment, there doesn’t appear to be a scaled LoRA for Qwen-Edit-2511, unlike Qwen-Edit-2509, which provides a proper FP8-compatible variant. Because of this, using the BF16 2511 Lightning 4-step LoRA with the FP8 2511 base model results in visible artifacts.

I also tried:
qwen_image_edit_2511_fp8_e4m3fn_scaled_lightning_comfyui.safetensors

However, the output quality is significantly worse compared to running:
qwen_image_edit_2511_fp8mixed.safetensors + BF16 2511 Lightning 4-step LoRA

Is there an official FP8-scaled (or otherwise FP8-compatible) 2511 Lightning 4-step LoRA planned? If not, what is the recommended setup for using Qwen-Edit-2511 Lightning with an FP8 base model without introducing artifacts?

[C0untFloyd]

Same here. Even the Q5 GGUF models have a significant better output to the fp8 version. I tried both, the _comfy version and the other.

[X-niper]

@menahem121

Please note that the Lightning LoRA has been merged into the qwen_image_edit_2511_fp8_e4m3fn_scaled_lightning_comfyui.safetensors model. It should be used as a standalone model for 4-step inference

[C0untFloyd]

@X-niper

Oh ok, this means we have to remove the lightning 4 steps lora from the workflow and instead just use this model (which also includes the relighting lora?) with 4 steps?
Thanks for the answer and merry xmas!

[menahem121]

@menahem121

Please note that the Lightning LoRA has been merged into the qwen_image_edit_2511_fp8_e4m3fn_scaled_lightning_comfyui.safetensors model. It should be used as a standalone model for 4-step inference

Obviously lol..

[menahem121]

@X-niper

Oh ok, this means we have to remove the lightning 4 steps lora from the workflow and instead just use this model (which also includes the relighting lora?) with 4 steps?
Thanks for the answer and merry xmas!

Yes it means you just need to load the model without any Lora's because the Lora is already there baked in.
Please do a comparaison of the model + Lora VS the all in one model and come back here 😉

[C0untFloyd]

Ok my bad I'm sorry 😶
It's working great now, thanks again. At least for me this issue can be closed.