Odyssey-2 Max

Starchild-1 is the first real-time multimodal world model, built to simulate both the visuals and sounds of the world in real time. Unlike language models, which learn from text, world models learn directly from the world itself through pixels, motion, and actions encoded in large-scale video, becoming capable of understanding and simulating an approximation of the world as it evolves. Starchild-1 goes beyond traditional world models, which have mostly focused on visual generation alone, by autoregressively generating synchronized audio and video while continuously responding to streaming user input. Instead of producing a fixed offline clip, it predicts the next audio and video state of a world based on past observations and live inputs, enabling environments, conversations, ambient sound, and world dynamics to change interactively. Users can stream text, speech, and action inputs into the model during rollout, dynamically altering what is seen and heard in real time.

Odyssey-2 Pro is a frontier general-purpose world model that generates continuous, interactive simulations you can integrate into products via the Odyssey API, marking a pivotal moment for world models similar to GPT-2 in language. It’s trained on large amounts of video and interaction data to learn how the world evolves frame-by-frame and outputs minutes-long simulations that can be interacted with in real time, not fixed short clips. Odyssey-2 Pro delivers improved physics, richer dynamics, more authentic behaviors, and sharper visuals by streaming 720p video at up to ~22 FPS that responds instantly to prompts and actions, and it supports embedding interactive streams, viewable streams, and parameterized simulations into applications with simple SDKs in JavaScript and Python. Developers can integrate the model with under ten lines of code to create open-ended, interactive video experiences where users’ inputs shape evolving scenes.