Odyssey publishes Starchild-1, a real-time audio-video world model

Odyssey releases Starchild-1, an autoregressive world model generating synchronized audio and video in real time based on past inputs and user actions.

Odyssey releases Starchild-1, an autoregressive world model that simultaneously generates sound and image in real time. Whereas classic world models are limited to visual synthesis, and audio-video systems like DeepMind's Veo produce fixed-duration clips offline, Starchild-1 continuously predicts the next audio and video state, based on past observations and user input.

Specifically, the system accepts text, voice, or actions injected during generation, which simultaneously modifies the displayed images and rendered sounds. Environments, conversations, soundscapes, and scene dynamics evolve with interaction rather than following a fixed trajectory.

The technical report details several new components. A causal distillation procedure converts a bidirectional audio-video model into a real-time compatible autoregressive system, while preserving the synchronization of both modalities. An asynchronous KV cache architecture and a rollout adaptation strategy absorb the differences in frequency and information density between sound and image, sensitive points for stabilizing long generations.

The lab frames this work within the perspective of general world intelligence and mentions applications in robotics, education, gaming, health, or defense. Starchild-1 is made available as a preview, accompanied by a technical report.