Abstract

This publication discloses an Asymmetric Mask-Guided Attention mechanism designed for video object removal in 3D Diffusion Transformers (DiTs). Standard attention mechanisms allow unrestricted bidirectional information flow, enabling erased object tokens to pollute unmasked background queries (mask noise bleed) and causing boundary discontinuities during iterative denoising. The disclosed technology introduces an asymmetric information gate paired with multi-resolution background key/value pooling. Background query tokens predominantly attend to background key tokens, isolating clean background regions from erased noise. Conversely, query tokens located inside erased regions attend to both interior hole keys and spatially average-pooled background keys. This asymmetric design preserves background fidelity, mitigates boundary seams, and substantially reduces computational complexity by up to ๐‘Ÿ(1 โˆ’ ๐‘Ÿ)(2 โ€“ 1/๐‘†2)ย relative to standard dense attention.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.

Share

COinS