World models
Viggle JST-2: a world model built in 3D, not pixels
Viggle's JST-2 generates an explorable 3D scene in about a second by modelling the world as persistent atoms rather than video frames. It also makes characters with motion built in, which matters for anyone building interactive worlds.
Most world models learn by predicting video. Viggle has taken a different route. On 19 October it released JST-2, which it calls a physics-native world foundation model, and which represents a scene as objects in 3D space rather than as a stream of frames.
What happened
Viggle, best known for its meme-video app, says it spent four years on the architecture behind JST-2. JST stands for Joint Space-and-Time. The first version, JST-1, powered the meme app. JST-2, developed between 2024 and 2026, extends the same idea to scenes, characters and motion in one model.
The model is available in the browser and through an API, including for use from coding agents.
How it works
JST models the world as persistent “atoms”. Each atom stands for a piece of matter or an object part and keeps its identity as it moves and deforms. It has an explicit xyz position, plus learned vectors for its attributes and dynamics. Viggle calls these Physics-Native Tokens.
An encoder turns images into tokens, a generator can produce them from a prompt, and a decoder turns them into images or 3D Gaussian splats. Because the splats render locally in the browser, a generated scene can be revisited from new angles without running the model again.
The numbers
Viggle reports that JST-2:
- generates a 3D scene in about one second of inference on an NVIDIA B200, at a cost below $0.01, against roughly five to six minutes for World Labs’ Marble in its own comparison;
- accepts up to 64 input views as a spatial context window, so a scene can be expanded one view at a time;
- creates a 3D character from a single image, with motion dynamics learned rather than rigged by hand;
- beat Meshy 7.1 and Tripo H3.1 in blind human evaluations of character generation (overall preference 59.9% and 66.9%) and animation readiness (70.6% and 89.6%).
These are Viggle’s own benchmarks, scored as win rate plus half the draw rate. We have not reproduced them.
Limits
Viggle says upcoming releases will add text prompts to guide scene expansion, more splats per scene for finer detail, and better clothing and hair movement. Today’s characters move through learned dynamics only. Interactions between objects still rely on conventional rule-based simulation. Viggle’s stated aim for JST-3 is to learn those too.
The scenes are static geometry you can explore. This is not yet a playable, rule-bound world.
Our take
Pixel-space models are impressive but hard to control, because the physical state of the world is only implied by how it looks. A 3D representation with identity-keeping objects is closer to what a simulation needs: things that persist, can be queried and can be changed.
For people building interactive worlds, the useful part is the pairing of fast scene generation with characters that arrive animation-ready. If those hold up outside demos, prototyping a place and its inhabitants gets much cheaper. The harder question, which JST-3 is meant to address, is behaviour. We explore why that is where stories actually grow in Living worlds: why simulation is the next storytelling medium.