🛰️ Daily AI Frontier
‹ back to 2026-08-13

LiveAnimate: Stable Long-Form Streaming Human Animation in Real-Time

arXiv cs.CV Multimodal & Generative Yuxuan Zhang, Haozhong Xiong, Yubo Huang, Jiayi Song, Jinpeng Yu, Haofan Wang, Jiaming Liu, Ruihua Huang, Liwei Wang 2026-08-12
Representative image for LiveAnimate: Stable Long-Form Streaming Human Animation in Real-Time

TL;DR - LiveAnimate is a 14B-parameter video diffusion system for stable, real-time, pose-driven human animation from a single reference image. It enables long-running interactive applications while keeping memory and latency constant.

  • Converts a bidirectional DiT into a block-causal generator, then distills sampling to three steps.
  • Uses Pose-Retrieval Sink Attention and a bounded KV cache to preserve appearance over long streams.
  • Achieves 19.63 FPS on two NVIDIA H100 GPUs using sequence parallelism and operator fusion.
  • Maintains nearly constant identity and perceptual quality across a three-minute benchmark.

view merged work →