LiveAnimate Streams Human Animation at 19.63 FPS on Two H100s

Diffusion pipelines have owned high-quality pose-driven human animation, but the price has always been offline wait times measured in minutes to hours per clip. A new arxiv preprint, LiveAnimate, claims to break that constraint at billion-parameter scale, reporting 19.63 FPS streaming inference from a 14B-parameter video Diffusion Transformer running on two NVIDIA H100 GPUs.

The authors describe a two-stage training pipeline. A Reference-Anchored Teacher-Forcing Adaptation turns a pretrained bidirectional DiT into a block-causal autoregressive generator, and Block-wise Self-Forcing Distillation then squeezes the sampling budget down to three steps. To keep the subject's appearance stable across long streams, they introduce Pose-Retrieval Sink Attention, a bounded KV-cache scheme combining a Static Sink that permanently anchors the first generated block, a Dynamic Sink that holds a pose-retrieved historical block, and a three-slot Rolling Window. The stated intent is that memory and per-block latency stay constant regardless of stream length. Ulysses sequence parallelism and operator fusion carry the rest.

If the numbers hold up, this reshapes economics for telepresence, live avatars, and virtual production, categories that have been stuck choosing between real-time GAN-quality output and diffusion-quality output rendered overnight. Two H100s per concurrent stream still puts this squarely in enterprise territory rather than on consumer laptops, but it shifts the frontier from "diffusion cannot stream" to "diffusion streams if you pay for the silicon."

Several things the abstract does not settle. There is no comparison against a named baseline with a specific quality metric such as FVD or identity score, no note on per-GPU memory footprint, and no word on whether weights or code will be released. Pose-Retrieval Sink Attention leans on the assumption that useful poses recur, which is fair for looping avatar motion but less obvious for arbitrary telepresence. Self-reported preprint benchmarks also tend to shrink once independent teams try to reproduce them.

The interesting question is which cloud vendor or avatar platform picks this architecture up first, because whoever does gets to sell diffusion-quality live video of a person as an API before the rest of the field catches up.