Streaming-Aware Diffusion for Real-Time Video Super-Resolution via Cross-Step Attention
This framework adapts single-image diffusion models for fast video super-resolution by using a streaming approach that exploits video's sequential nature. It introduces Cross-Step Attention to reuse features across frames and steps, and Trajectory-Coupled Diffusion Scheduling to align states. This results in a streaming pipeline that reduces computation from O(N*S) to O(N+S), achieving real-time performance over 40 FPS.