Core Focus: Video-level object removal. Given a sequence of video frames and a corresponding mask, it seamlessly removes the masked object and fills in the background while maintaining temporal consistency, minimizing artifacts or remnants.

Method Highlights:

Minimum-Maximum Optimization: Tames bad noise during training and inference, improving the model's robustness to masked regions and reducing the probability of object regeneration.

Two-Stage Architecture: First, a simplified DiT (Diffusion Transformer) structure is used to learn the removal capability; then, a version with fewer sampling steps and faster inference is obtained through "CFG de-distillation."

Efficiency Features: Extremely low inference steps (approximately 6 steps in the official example), and does not rely on CFG, resulting in high speed and low resource consumption, suitable for long videos/batch processing. References

Wan2.1 Minimax-Remover - Video erase -KJ

Rating & Review