Question about the masking strategy

[fermigo66-collab]

Hello Authors, thanks for the great work!

In the tech report there is an interesting masking strategy for edit pre-training, would you like to share its implementation or more details about this strategy? Is it also possible to use existing segmentation tools for the purpose?

[Carlofkl]

Hello Authors, thanks for the great work!

In the tech report there is an interesting masking strategy for edit pre-training, would you like to share its implementation or more details about this strategy? Is it also possible to use existing segmentation tools for the purpose?

Hi, thanks for your interest!

Our masking strategy is based on a simple pixel-level difference between the source and target images, followed by dilation and filtering. More details can be seen in UniWorld-V1: https://arxiv.org/abs/2506.03147. It's lightweight and requires no external models

Regarding using existing segmentation tools (e.g., SAM): yes, it's feasible and could provide more precise masks. However, we found that pixel-diff based masking is sufficient for our training objective and is significantly faster, especially when processing millions of training pairs. Segmentation-based approaches would introduce considerable computational overhead at scale.