Does EditScore support multi-images to image (virtual tryon, multi-concept fusion)?

[robertchen245]

Does EditScore support multi-images to image, or can EditScore take multiple input images?

[sabulin]

EditScore is currently trained as a reward model for single-image editing, since its training data only contains single-image editing examples.

Interestingly, after RL training with EditScore on single-image editing tasks, we observed improved performance of the policy model on the OmniContext benchmark (which involves multi-image editing). This suggests that the reward signal learned from single-image editing can generalize to some multi-image editing scenarios.