minWM: A Full-Stack Open-Source Framework for Real-Time Interactive Video World Models

Researchers released minWM, an open-source framework for turning video diffusion models into interactive video world models.

The system fine-tunes bidirectional text-to-video and image-to-video models, then uses causal training and distillation to enable few-step autoregressive rollout.

Built on models like Wan2.1-T2V-1.3B, HY1.5-TI2V-8B, and HY-WorldPlay, it targets low-latency inference and camera control.

The release provides a reproducible pipeline for real-time, camera-controllable video generation and world modeling.