FlashVSR+ runs untiled 4K video super-resolution on one RTX 4090 with 2x speedup
MMLab-CUHK · hf · 2026-10-14
Researchers from Shanghai AI Lab and CUHK MMLab present FlashVSR+, tackling costly high-res inference and preference misalignment in video super-resolution.
- Architecture: hybrid attention exploiting strong locality with sparse global interaction, delivering up to 2x DiT speedup and 45% lower peak memory — enabling spatially untiled 4K VSR on a single 24GB RTX 4090
- Training: Reward-Yield Distribution Matching Distillation (Ry-DMD) injects black-box rewards via reward-weighted teacher rollouts without backprop through the reward model, and generalizes to T2I
- Consistent user-preference gains; code, models, and training data released at flashvsr.tech
More from Research
- Swaayatt Demos On-the-Fly Camera-LiDAR Calibration Without Markers in the Wild — sanjeevs_iitr · 2026-10-14
- LongR: RL with dense utility rewards boosts long-context reasoning by 9% on LongBench v2 — PKU · 2026-10-14
- On-policy distillation study: 4 prompts match 3,840, but prompt utility depends on the teacher-student pair — Tencent-Hunyuan · 2026-10-14
- RISC-V crypto accelerator SoC gets full place-and-route done with ~$2,000 of LLM tokens — IgorCarron · 2026-10-14
- Tufts and Santa Fe Institute paper models LLM adoption as a 'cognitive virus' — alex_verem · 2026-10-14
- PyVBBC 1.5 Released After 3 Years, Bringing Major Upgrades to Bayesian Model Inference — twiecki · 2026-10-14