RealtimeWAM: One-Step Asynchronous World Action Models
RealtimeWAM enables one-step action generation and asynchronous inference, with about 25x speedup and under 1% drop.
World Action Models reuse video representations for action prediction, but multi-step denoising and sequential expert execution slow inference. RealtimeWAM uses Teacher-Anchored Consistency Distillation for accurate one-step actions and Cross-Expert Wavefront Pipelining to overlap the experts via block-wise KV-cache sharing. On LIBERO, LIBERO-Plus, and RoboTwin, variants including Fast-WAM and Faster-WAM lose under 1% performance while reaching about 25x end-to-end speedup on an H100. Code and checkpoints are released.
- TACD supervises one-step actions with the frozen teacher's multi-step rollout endpoint.
- CEWP overlaps video and action experts through block-wise video KV-cache sharing.
- Benchmarks show under 1% drop with about 25x end-to-end speedup on H100.
Full article210 words · extracted from huggingface.co · click to collapse
World Action Models (WAMs) incorporate visual representations from video generation backbones to guide action prediction. Recent efficient WAMs adopt Mixture-of-Transformers (MoT) architectures and compute video representations once for reuse by the action expert. However, intra-expert iteration (\ie, multi-step action denoising) and inter-expert waiting (\ie, sequential execution of the video and action experts) still limit inference efficiency. To this end, we present RealtimeWAM, an extremely efficient WAM variant with one-step action generation and asynchronous inference, addressing these two bottlenecks. To reduce intra-expert iteration, we propose Teacher-Anchored Consistency Distillation (TACD) to address a local-global error gap: low local consistency error alone does not guarantee accurate final actions. TACD supplements local consistency with explicit supervision from the frozen teacher's multi-step rollout endpoint, enabling accurate one-step action generation. Additionally, we propose Cross-Expert Wavefront Pipelining (CEWP) to eliminate unnecessary expert-level waiting. It overlaps the two experts through block-wise sharing of the video KV cache, synchronizing only immediately before the corresponding action attention consumes it. Extensive experiments across diverse benchmarks (\eg, LIBERO, LIBERO-Plus and RoboTwin) and model variants (\eg, Fast-WAM and Faster-WAM) demonstrate the superiority of RealtimeWAM. Notably, RealtimeWAM maintains near-lossless performance (\ie, <1% drop) across these benchmarks while delivering significant end-to-end speedup (\eg, sim25times on H100). Our code and checkpoints are available via this https://github.com/ModelTC/LightX2V/tree/main/examples/realtimewam{link}.
Text extracted automatically; images, tables and formatting may be missing. Original: https://huggingface.co/papers/2610.06617