Allow training Qwen Edit Plus without control images #682
Description
Qwen Edit 2509 can be used both as image-to-image, and text-to-image model. It doesn't need to have an input image.
I want to train a style lora that will work for both cases (I2I and T2I). So my datasets for this lora are split. Dataset A has images with captions. Dataset B has image pairs with edit instructions.
When I try to train with such a dataset (with --edit_plus flag), I get error: AssertionError: Item must have control content for Qwen-Image-Edit for images from dataset A.
I understand that text encoder works differently with control images present. It should look at context: if there's no control image specified, encode as in the non-edit_plus case. For UX sake, an extra flag can be provided in dataset.toml: allow_no_control = true so users can't accidentally mess up their datasets.
If there's a workaround right now, I would like to know about it