Flash-WAM: Modality-Aware Distillation for World Action Models
World-action models (WAMs) use iterative diffusion to generate video and robot actions but are too slow for real-time control, requiring many denoising steps. Flash-WAM solves this by using a modality-aware distillation framework that selects different mathematical functions for video and action streams based on their unique noise patterns. This allows inference in just one step per modality, achieving real-time performance while maintaining high task success rates.