LynnReal-Hand: hand-controlled video world model hits 18 FPS on one H200
Shaohao Rui · hf · 2026-11-17
LynnReal AI released a technical report for LynnReal-Hand, a hand-controlled video world model enabling direct object interaction via hand motions, with datasets, weights and full training/inference code to be open-sourced.
Approach
- 100K curated interaction clips across real and synthetic environments, with hand-aware supervision for fine-grained hand-object dynamics
- Causal flow map for few-step generation; region-aware score distillation blending video and image priors
- Ultra-Fast-VAE, a compact video decoder cutting compute while preserving reconstruction fidelity
Results
- Full-sequence model: FVD 181.67, MANO Dice 93.02; 3-step causal model: MANO Dice 92.49
- Real-time 540p streaming on a single H200 at 18.03 FPS, 741ms latency, 34.89GB peak VRAM
Outperforms prior methods on HandBench in hand-following accuracy and visual quality; weights are live on Hugging Face.
More from Embodied
- Two Teslas With Identical FSD Deadlock for Minutes on a Narrow Mountain Road — aakashgupta · 2026-11-17
- Koop CEO on robot insurance: rating each humanoid and charging by the hour, like API usage — audrow · 2026-11-17
- YC-Backed Darwin Robotics Hires Founding Team for Data Center Robot Technicians — Kyrannio · 2026-11-17
- Wayve's Marc Wulfmeier: truly useful household humanoids won't arrive before 2050 — m_wulfmeier · 2026-11-17
- 8th NeurIPS Robot Learning Workshop to debate: Is Physical AI truly going zero-shot? — m_wulfmeier · 2026-11-17
- VR human demos fix persistent failures in point-cloud grasping policies — hcltech-robotics · 2026-11-17