Feature request: Multi-image edit / multi-control training support for FLUX.2 Klein and Qwen-Image-Edit

[Xabsurd]

Hi, first of all, thank you for maintaining diffusion-pipe.

This is currently my favorite trainer. I really like its stability, DeepSpeed support, dataset handling, and the way it supports larger models for long training runs.

I would like to ask whether multi-image edit training / multi-control training could be supported in the future, especially for models such as:

• FLUX.2 Klein edit
• Qwen-Image-Edit
• other image-editing models that support multiple reference images

At the moment, the edit training workflow seems to support a paired dataset like:

[[directory]]
path = "/path/to/target"
control_path = "/path/to/control"

This works well for single-image editing, where each target image has one corresponding control/reference image.

However, for multi-image editing tasks, it would be very useful to support multiple reference images for one target image. For example:

target image: final edited result
control image 1: person reference
control image 2: clothing reference
control image 3: background / style reference
caption: "Use the person from image 1, the outfit from image 2, and the background from image 3."

A possible dataset format could be something like:

[[directory]]
path = "/path/to/target"
control_paths = [
  "/path/to/ref_person",
  "/path/to/ref_clothing",
  "/path/to/ref_background"
]

[Xabsurd]

This PR should resolve this issue. #529

It adds control_paths = [...] support for edit datasets, so a single target sample can be conditioned on multiple control/reference images while keeping the existing single control_path behavior unchanged.

I also validated it with a small Flux2 Klein 9B multi-control edit training run:

• 10 real edit samples
• 2 control images per target
• 512 resolution
• 5 repeats per epoch
• 20 epochs
• LoRA generation comparison with and without the trained adapter

Training completed successfully and the LoRA produced the expected edited output.