Flash-WAM: Modality-Aware Distillation for World Action Models
Researchers have developed Flash-WAM, a new framework for world-action models that significantly speeds up inference time. Traditional models require many denoising steps, making real-time control difficult. Flash-WAM employs a modality-aware step-distillation technique, adapting to the distinct noise characteristics of video and action streams. This allows for a single-step inference process, reducing latency from over 8 seconds to under 350 milliseconds on NVIDIA L40S hardware, a 23x improvement. AI
IMPACT Enables real-time robotic control and manipulation by drastically reducing inference latency for world-action models.