Overworld releases Waypoint-1, a real-time video diffusion model controlled by text, mouse, and keyboard
- Overworld released weights for Waypoint-1-Small, its interactive video diffusion model; Waypoint-1-Medium is listed as coming soon.
- Waypoint-1 generates each frame with text prompts, mouse movement, and any keyboard input as context, and Overworld says users can supply frames to create an interactive world.
- The model uses a frame-causal rectified flow transformer trained on 10,000 hours of video game footage paired with control inputs and text captions.
- Waypoint-1 is a latent model trained on compressed frames and pre-trained with diffusion forcing, which denoises future frames from past frames under a causal attention mask.
- Overworld post-trains the model with self forcing because randomly noising frames during diffusion forcing can cause inference mismatch, error accumulation, and noisy long rollouts.