Overworld has released Waypoint-1, a real-time interactive video diffusion model. It generates and streams procedural video worlds controllable via text, mouse, and keyboard. Unlike previous models that fine-tune pre-trained systems, Waypoint-1 was trained from scratch on 10,000 hours of game footage paired with control inputs, aiming for zero-latency interaction.

The model uses a frame-causal rectified flow transformer architecture. Training employs diffusion forcing and self-forcing techniques to reduce error accumulation during long rollouts. Overworld also open-sourced its inference library, WorldEngine, which runs efficiently on consumer hardware like a 5090 GPU, sustaining high frame rates for seamless streaming.

This approach targets a key limitation in world models: laggy, simplistic controls. By training for interactivity from the start, Waypoint-1 aims to create more responsive and immersive experiences. The upcoming hackathon on January 20, 2026, with a 5090 GPU prize, will test its extensibility and attract developer interest in building new interactive applications.