[Feat] support JoyAIImagePipeline

[RuixiangMa]

Support JoyAIImagePipeline

• https://github.com/jd-opensource/JoyAI-Image

• jdopensource/JoyAI-Image-Edit

Prompt	image
NO
add hat in the dog
Rotate the dog to show the left side view
Rotate the dog to show the right side view

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline?
• Did you read our philosophy doc (important for complex PRs)?
• Was this discussed/approved via a GitHub issue or the forum? Please add a link to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

[yiyixuxu]

thanks for the PR, i left some initial comments

[yiyixuxu]

can you not import from wan? let's use #Copied from isntead

[RuixiangMa]

I will remove the direct imports from Wan and inlined the reused helper blocks in JoyAI with # Copied from comments instead.

[yiyixuxu]

what's going on here? is there any reason we cannot make eit work with DiffusionPipeline.from_pretrained?

[RuixiangMa]

The upstream JoyAI checkpoint is not a standard diffusers repo layout, so it cannot be loaded by DiffusionPipeline.from_pretrained directly.

I agree the better long-term solution is to add a one-time conversion/export to a standard diffusers checkpoint and then rely on the base from_pretrained path for the exported directory

[yiyixuxu]

@claude
can you tell me what's the difference between this and FlowMatchDiscreteSchedule?

[RuixiangMa]

I double-checked this. It is close to FlowMatchEulerDiscreteScheduler, but not numerically identical as-is because JoyAI uses its own timestep/sigma shifting behavior.

I will refactorit to subclass FlowMatchEulerDiscreteScheduler and only keep the JoyAI-specific differences in the scheduler implementation.

[RuixiangMa]

@yiyixuxu Thanks for the comments! PTAL

[Moran232]

@RuixiangMa
Thank you very much for your attention and your pull request!

The JoyAI team is the AI team from JD Explore Academy, and this is our first open-source release of a self-developed diffusion model.

After open-sourcing the model on GitHub, we spent one week adding support for the diffusers library. The official name of the model is JoyAI-Image-Edit. In the diffusers pipeline, we would like to name it JoyImageEditPipeline, as we plan to release additional models soon—including JoyImagePipeline, JoyVideo, and others to form a cohesive model family. Therefore, naming this PR JoyAIImagePipeline might not be the most appropriate choice.

I have submitted a new Pull Request that includes the following key updates:

• The model now fully adopts WanVAE, completely reusing WanVAE's existing code.
• Removed the dependency on FlashAttention.
• Fixed discrepancies between the scheduler implementation and the originally open-sourced version.

To facilitate future model updates and open-source releases, could you please migrate your code from this PR to my newly submitted PR? You’re also welcome to directly add any other valuable suggestions or changes to that PR.