Performance Improve for Qwen Image Edit

[naykun]

We observe that conditioning resolution strongly influences the alignment of source and target images. This PR introduces a unified resize logic, resulting in significantly enhanced editing performance.

Please kindly give it a review.
cc @yiyixuxu @a-r-r-o-w @sayakpaul

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[linoytsaban]

thanks!

since _auto_resize is removed from the preprocessing, should it be removed altogether, here as well?

diffusers/src/diffusers/pipelines/qwenimage/pipeline_qwenimage_edit.py

Line 549 in 70f9ea9

_auto_resize: bool = True,

[naykun]

thanks!

since _auto_resize is removed from the preprocessing, should it be removed altogether, here as well?

diffusers/src/diffusers/pipelines/qwenimage/pipeline_qwenimage_edit.py

Line 549 in 70f9ea9

_auto_resize: bool = True,

Sure, resolved in latest commit.

[asomoza]

thanks!

[gluttony-10]

We observe that conditioning resolution strongly influences the alignment of source and target images. This PR introduces a unified resize logic, resulting in significantly enhanced editing performance.

Please kindly give it a review. cc @yiyixuxu @a-r-r-o-w @sayakpaul

This is an incorrect update. Picture consistency is lost again.

[naykun]

We observe that conditioning resolution strongly influences the alignment of source and target images. This PR introduces a unified resize logic, resulting in significantly enhanced editing performance.
Please kindly give it a review. cc @yiyixuxu @a-r-r-o-w @sayakpaul

This is an incorrect update. Picture consistency is lost again.

Could you please share your img, prompt and your experiment setting?

[gluttony-10]

#12188 Its right. #12190

[naykun]

@gluttony-10 I noticed the two cases use different prompts, and the seeds are not the same—this could lead to different outputs. Also, are you using prompt enhancement? If possible, please share your input image, and I’ll run a local test for comparison.

[gluttony-10]

@gluttony-10 I noticed the two cases use different prompts, and the seeds are not the same—this could lead to different outputs. Also, are you using prompt enhancement? If possible, please share your input image, and I’ll run a local test for comparison.

Sincere apology, it's a problem with the prompt words. Without tags that remain unchanged, there is a significant difference in consistency.