Dual-PAG: Bidirectional Perturbed-Attention Guidance for Residue-Free Video Editing

Text-driven video editing aims to modify desired semantic content while preserving the motion, layout, and temporal coherence of a source video. However, velocity-based editing methods often leave source residue when the target edit requires a large object-level transformation. To address this issue, we propose Dual-PAG, a bidirectional Perturbed Attention Guidance framework that suppresses source-side structural traces while reinforcing target-side geometry. Dual-PAG further uses cross-attention-based masking to localize the correction to the edit region. Dual-PAG preserves the single-trajectory nature of inversion-free video editing while more effectively reducing source residue in challenging object replacement scenarios.