FlashVSR+ introduces a hybrid attention mechanism that combines strong local correlation from pretrained VSR models with sparse global interaction, achieving up to twice the speed of DiT and 45% lower peak memory, which permits spatially untiled 4K video super‑resolution on a single 24 GB RTX 4090. The architecture retains one‑step generation via diffusion distillation while reducing inference cost. For training, FlashVSR+ employs Reward‑Yield Distribution Matching Distillation (Ry‑DMD), a method that weights teacher rollouts by an external reward signal and matches the resulting distribution without backpropagating through the reward model, thereby aligning one‑step generation with human preferences. The hybrid attention reduces computational complexity by leveraging locality while preserving long‑range consistency, and the Ry‑DMD loss integrates reward signals from off‑the‑shelf evaluators, enabling preference‑aligned training without modifying the reward network. Experiments on standard VSR benchmarks demonstrate a 0.12–0.18 dB gain in PSNR over baseline distilled models, and the one‑step generation eliminates iterative sampling, further accelerating inference. Ry‑DMD also transfers to image‑to‑image (T2I) generation, demonstrating broader applicability. The authors release the code, pretrained models, and training data at the provided URL.