Blur images when doing text2img with the qwen-image-edit loras

[menahem121]

I'm using ComfyUI v0.3.52 with an L40s GPU and the FP8 version of the Qwen-Image-Edit model.

I've tried:
Qwen-Image-Edit-Lightning-4steps-V1.0-bf16.safetensors
Qwen-Image-Edit-Lightning-4steps-V1.0.safetensors

Both give blurry results.
However, Qwen-Image-Lightning-4steps-V1.0.safetensors works very well with Qwen-Image-Edit. 🤷‍♂️

From comparing Qwen-Image-Lightning-4steps-V1.0.safetensors and Qwen-Image-Edit-Lightning-4steps-V1.0.safetensors, my conclusion is that the Qwen-Image-Lightning-4steps-V1.0.safetensor lora preserves the input image more accurately.

So I wonder: does this new edit LoRA really make sense?

[X-niper]

@menahem121

Hi,

Q1:

The new Edit-lora weights are trained with Qwen-Image-Edit on editing tasks and are not expected to work on T2I tasks. However, reversely, the weights trained on T2I tasks work on Edit tasks. Similaryly, we also observe the lora weights trained on T2V work on I2V, too.

Q2:

We observe the weights trained on the same task as the test task should work better.

Could you please share us the test images and prompt, so that we can find out the limitation of Qwen-Image-Edit-Lightning-4steps-V1.0.safetensors?

[menahem121]

@menahem121

Hi,

Q1:

The new Edit-lora weights are trained with Qwen-Image-Edit on editing tasks and are not expected to work on T2I tasks. However, reversely, the weights trained on T2I tasks work on Edit tasks. Similaryly, we also observe the lora weights trained on T2V work on I2V, tool.

Q2:

We observe the weights trained on the same task as the test task should work better.

Could you please share us the test images and prompt, so that we can find out the limitation of Qwen-Image-Edit-Lightning-4steps-V1.0.safetensors?

Sorry for the late answer. I wanted to do more tests, and I got to the conclusion that you are right. The qwen-image-edit Lora is working better for inpainting and img2img than the qwen-image one.

But I still maintain that it's a pain to have to use both Loras for txt2img and editing because VRAM is not infinite, and loading again each Lora takes time. Also, this means that the qwen-image-edit Lora can't be baked into finetune models because it will make it unable to do txt2img.

[yyfcc17]

@X-niper hello, i have a question.

how do you obtain the editing task data to distill the editing model?

i think the dataset is the main obstacle to distill the editing model.

thank you.