Viggle's JST-2 builds explorable 3D worlds from a single prompt

Many people remember Viggle as the viral meme app that let anyone stick celebrities and friends into hilarious dance clips. But the team behind those 50 million users just revealed what they have actually been building for the past four years.

Viggle introduced JST-2, a foundation world model designed from scratch around physical space. Instead of hallucinating flat video frames, JST-2 converts simple text prompts or a few images into fully explorable 3D environments in about one second, at an inference cost under a single cent.

Even better, the 3D characters it generates come with motion built in, completely skipping the painful manual rigging step that has slowed down digital artists for decades.

Atoms over pixels

Most current world models try to learn physics purely in pixel space. They generate stunning visuals, but the moment a camera turns away and looks back, objects morph or disappear entirely because the model never truly knew where anything was.

JST stands for Joint Space-and-Time. Instead of predicting pixels, Viggle built a physics-native architecture that treats the world as persistent 3D tokens representing actual matter.

Each physical token keeps its own xyz coordinates alongside learned vectors that describe its material traits and dynamics. In simple terms, the model tracks physical matter rather than pictures of matter.

Because objects maintain their identities and locations, a scene stays completely consistent as you move through it. Generated worlds decode directly into 3D Gaussian splats, allowing users to inspect and wander through environments in real time right inside a standard web browser without constantly rerunning heavy cloud inference.

It also supports a spatial context window of up to 64 views, letting you expand a room or an open landscape piece by piece into a cohesive map.

Rig-free characters and cleaner motion

Anyone who has touched 3D animation knows that modeling is only half the battle. Rigging skeletons and painting skin weights to prevent weird mesh tearing during movement is notoriously tedious.

JST-2 sidesteps this by baking deformation dynamics straight into the tokens. The model already knows how skin, clothing, and limbs fold during movement, allowing it to turn even a cropped or partially occluded image into an animatable 3D character immediately.

In blind evaluation tests against leading 3D generators Meshy 7.1 and Tripo H3.1, JST-2 came out on top in character fidelity, especially in facial sharpness and hand articulation.

When characters move, the deformation across body parts and clothing feels remarkably natural. Furthermore, its generated full-body motions can map onto standard character rigs or serve as rich training grounds for autonomous robots practicing spatial navigation.

Viggle has not open-sourced the underlying weights, but the tool is available to try for free directly in the browser.

Why it matters

We think JST-2 is a major reality check for the generative AI race. It proves that teaching a model spatial geometry directly beats trying to brute-force 3D consistency out of 2D video diffusion.

Delivering an explorable 3D scene in one second for under a penny fundamentally changes the math for game prototyping, interactive virtual worlds, and architectural previsualization. For indie developers and solo animators, wiping out the rigging barrier turns character creation into an almost casual workflow.

There are clear limits today. Fine hair movement and highly complex, layered clothing simulations still need polishing, and keeping the model proprietary limits deep integration into local studio pipelines.

Still, Viggle has made a remarkably smart transition from a lighthearted meme factory into a serious frontier model lab. By modeling persistent physical space instead of fleeting pixels, JST-2 sets a practical benchmark for where interactive AI needs to go.