CanvasMAR: Improving Masked Autoregressive Video Prediction With Canvas

Researchers have developed CanvasMAR, a new masked autoregressive video prediction model that generates high-quality videos with fewer sampling steps by using a "canvas" approach that provides global structure early in the generation process. The model demonstrates superior performance on major benchmarks including BAIR, UCF-101, and Kinetics-600, rivaling advanced diffusion-based methods.

• →CanvasMAR introduces a canvas-based approach that provides global structure to improve video frame synthesis quality.
• →The model uses a motion-aware sampling strategy that processes stationary regions before dynamic areas for better stability.
• →CanvasMAR achieves remarkable performance on Kinetics-600 dataset, competing with state-of-the-art diffusion models.
• →The approach significantly reduces the number of autoregressive steps needed for high-quality video generation.
• →Compositional classifier-free guidance is integrated to enhance both canvas and temporal conditioning.