FlashVSR+: Efficient Video Super-Resolution via Hybrid Attention and Reward-Yield Distribution Matching
FlashVSR+ claims untiled 4K video super-resolution on a single 24GB RTX 4090, with faster and lighter high-resolution inference.
The paper targets one-step VSR models distilled from diffusion systems, where high-resolution runs are still expensive and preference alignment is hard. Its hybrid attention keeps local computation dominant while preserving sparse global interaction, which the authors say gives up to a 2x DiT speedup and 45% lower peak memory. The training method, Ry-DMD, uses reward-weighted teacher rollouts to fold black-box rewards into distribution matching without backpropagating through the reward model. The authors also say the method improves user preference and generalizes beyond VSR to text-to-image generation. HF Daily Papers' note
The paper targets one-step VSR models distilled from diffusion systems, where high-resolution runs are still expensive and preference alignment is hard. Its hybrid attention keeps local computation dominant while preserving sparse global interaction, which the authors say gives up to a 2x DiT speedup and 45% lower peak memory. The training method, Ry-DMD, uses reward-weighted teacher rollouts to fold black-box rewards into distribution matching without backpropagating through the reward model. The authors also say the method improves user preference and generalizes beyond VSR to text-to-image generation. HF Daily Papers' note
score 4