Shanghai AI Lab's FlashVSR+ runs untiled 4K VSR on a 24GB 4090

FlashVSR+, a one-step video super-resolution framework from Shanghai AI Lab, MMLAB CUHK, Tsinghua University and CPII under InnoHK, pushes spatially untiled 4K VSR onto a single 24GB RTX 4090 while reporting up to 2× DiT speedup and 45% lower peak memory. The paper on Hugging Face credits a hybrid attention design that replaces most dense attention blocks with efficient local spatiotemporal attention and keeps a small number of sparse global blocks for long-range consistency.

The authors argue the shortcut is justified by what they see inside pretrained VSR backbones. "Attention is highly concentrated within compact neighborhoods in many blocks, suggesting that dense attention is unnecessary throughout the backbone, while fully local attention would be overly restrictive," they write. In their own phrasing the architecture "nearly halves peak memory," which is what lets 4K inference skip the spatial tiling that "limits global context and may introduce boundary artifacts" on consumer GPUs.

On the training side, the team pairs the new architecture with Reward-Yield Distribution Matching Distillation (Ry-DMD), which "uses reward-weighted teacher rollouts to incorporate black-box rewards into distribution matching, without backpropagating through the reward model." They validate it on OCR-guided VSR and report that it generalizes to text-to-image generation.

Code, models and training data are released at www.flashvsr.tech. The release lands amid a steady run of efficiency work in our AI-video tracker, where this is one of 53 items in the last 90 days. The abstract publishes no PSNR or LPIPS numbers against prior VSR baselines.