[Feature Request] Add FLUX.2 inpaint support

[firefighter-eric]

Hi, thanks for the great work on DiffSynth-Studio.

FLUX.2 is already supported for inference and training in the repository, and there is also support for image editing through edit_image in the FLUX.2 pipeline/examples. However, it seems that FLUX.2 still does not support a standard inpainting workflow with an explicit mask input.

It would be very helpful if FLUX.2 inpaint support could be added, for example:

Requested feature

Support masked inpainting for FLUX.2 models (such as FLUX.2-dev / FLUX.2-klein), with an interface similar to:

• image / edit_image
• mask_image
• prompt-guided masked region editing
• keeping unmasked regions unchanged as much as possible

Why this is useful

Many practical editing tasks need localized modification instead of full-image editing, for example:

• replacing or removing local objects
• changing clothing / accessories only in a selected region
• background cleanup
• PPT / poster / design asset partial editing

Right now, edit_image is useful for general image editing, but in many cases we need stricter spatial control through a mask.

Suggested scope

It would be great if the project could provide:

1. Inference support in Flux2ImagePipeline
2. An example script under examples/flux2/model_inference/
3. Documentation for the expected mask format and recommended preprocessing
4. (Optional) training / LoRA examples for inpaint-style tasks in the future

Example expected usage

image = pipe(
    prompt="replace the masked area with a red mug",
    edit_image=input_image,
    mask_image=mask,
    num_inference_steps=30,
    seed=0,
)

---

非常感谢 DiffSynth-Studio 的工作！

目前仓库已经支持 FLUX.2 的推理和训练，也看到 FLUX.2 有基于 edit_image 的图片编辑能力；不过看起来暂时还没有提供标准的 inpaint（掩码局部重绘） 支持。

希望可以为 FLUX.2 增加 inpaint 能力，例如支持：

• edit_image
• mask_image
• 基于 prompt 的局部区域编辑
• 尽量保持非 mask 区域不变

需求原因

很多实际场景都需要“局部可控编辑”，而不是整图编辑，比如：

• 替换某个局部物体
• 只修改人物衣服/配饰
• 局部背景清理
• 海报/PPT/design asset 的局部修补

edit_image 对整体编辑已经很有帮助，但很多场景还是需要 mask 来提供更强的空间约束。

建议支持内容

1. 在 Flux2ImagePipeline 中增加 inpaint 推理接口
2. 在 examples/flux2/model_inference/ 下提供示例
3. 文档说明 mask 格式、尺寸要求和预处理建议
4. （可选）后续补充 inpaint 相关训练 / LoRA 示例

期望的调用方式

image = pipe(
    prompt="replace the masked area with a red mug",
    edit_image=input_image,
    mask_image=mask,
    num_inference_steps=30,
    seed=0,
)