Viggle, the Toronto startup best known for the app that puts users' faces on dancing bodies, has launched JST-2, a model the company calls the first full-stack, physics-native world foundation model. The difference from traditional video generators lies in where it starts: rather than learning from pixels, the model builds its entire architecture directly in 3D.
According to Viggle, JST-2 generates a 3D scene in seconds, can expand it in real time, and creates characters with built-in motion. Environments are born with physics integrated, so objects and characters behave according to physical rules rather than being simply sequences of images.
The release marks a shift for the company. JST-1, introduced in 2024, was a video-3D model with an understanding of physics, and it powered the character swap videos that made Viggle known. Those results were video. With JST-2, the research leaves the consumer app for the first time and is made available to developers too. The model can be used in the browser through PINOC and via API, including from coding agents.
The same model can also reconstruct a full 3D environment from sparse real-world photos, a capability Viggle points to for simulation, mapping and other spatial applications. The Toronto team, led by co-founder and CEO Hang Chu, says it developed JST-2 between 2024 and 2026.
Image: viggle.ai
For people working on games and interactive environments, the shift is concrete: a model that produces manipulable geometry and scenes, not just footage, is closer to a production tool. The launch follows other announcements in the same area, such as GPT-6 Astra turning real videos into interactive 3D worlds and Tencent's WorldClaw, which generates editable 3D worlds from a prompt.
Whether JST-2 holds up outside the official demos remains to be seen: the launch figures come from the company itself and have not yet been checked by independent tests.
No comments yet. Be the first!