LynnReal-Hand: Unlocking World Interaction with Hand Control
Hand-controlled video generation reaches 18.03 FPS at 540p on a single H200, with 741ms latency.
LynnReal-Hand translates hand movements into generated video of object interactions. The researchers assembled 100,000 real and synthetic interaction clips, then combined hand-aware training with few-step generation and a compact video decoder. On HandBench, they report better hand-following accuracy and visual quality than existing approaches, with the three-step model scoring 92.49 on MANO Dice. They plan to release the datasets, model weights, and training and inference code.
HF Daily Papers' note
LynnReal-Hand translates hand movements into generated video of object interactions. The researchers assembled 100,000 real and synthetic interaction clips, then combined hand-aware training with few-step generation and a compact video decoder. On HandBench, they report better hand-following accuracy and visual quality than existing approaches, with the three-step model scoring 92.49 on MANO Dice. They plan to release the datasets, model weights, and training and inference code.
HF Daily Papers' note
score 5