This paper introduces Reward Forcing, a novel framework for efficient streaming video generation that tackles issues like diminished motion dynamics and over-reliance on initial frames. It achieves state-of-the-art performance by combining EMA-Sink for improved long-term context and Rewarded Distribution Matching Distillation (Re-DMD) to enhance motion quality.

Pyramid Sparse Attention (PSA) introduces multi-level pooled key-value representations to address the quadratic complexity and information loss challenges of existing sparse attention mechanisms. By dynamically allocating finer pooling levels to critical blocks and coarser levels to less important ones, PSA significantly expands the receptive field and preserves contextual information, outperforming current baselines in video understanding and generation tasks.