Summary
WAMs generate future video and robot actions through iterative diffusion requiring tens of denoising steps, which precludes real-time control. Flash-WAM identifies that off-the-shelf consistency distillation fails in the joint video-action setting because video and action streams use different SNR-shifted noise schedules — a fundamental asymmetry single-modality distillation cannot handle. Flash-WAM introduces modality-aware step distillation: a linear-gradient-scaling parametrization for the action stream’s low-noise regime, paired with a variance-preserving parametrization for the video stream’s high-noise regime. The result is a few-step WAM that maintains task performance while dramatically reducing inference cost.
Key Contributions
- Identifies the modality-SNR asymmetry problem preventing standard consistency distillation from working in joint video-action models
- Linear-gradient-scaling parametrization for action streams (low-noise regime)
- Variance-preserving parametrization for video streams (high-noise regime)
- Enables practical real-time WAM control with negligible task performance loss
Significance
Flash-WAM is the first principled solution to the speed bottleneck of WAMs, enabling the joint video-action diffusion paradigm to run at interactive rates — a prerequisite for real-world deployment.