Viggle has released JST 2, a world model that builds 3D characters, motion and scenes out of Gaussian splats.
Viggle's earlier model, JST-1, validated the company's architecture on character generation and animation and powered the meme-making app it became known for, which Viggle says has passed 50 million users. Those results were video. JST 2, developed from 2024 to 2026, extends the same architecture to scenes, characters and motion that share one 3D world.
JST stands for Joint Space and Time. In its research post, Viggle outlines a model that represents the world as persistent atoms in 3D space, each standing for a piece of matter or part of an object and keeping its identity as it moves and deforms. Every atom carries an explicit xyz position plus learned vectors for its attributes and dynamics, known as Physics-Native Tokens. An encoder infers those tokens from real images, a generator produces them from a prompt, and a decoder turns them into images or 3D Gaussian splats. Once a scene is decoded, it can be explored from any viewpoint in the browser without rerunning the model.
Viggle says JST 2 generates a 3D scene in about one second on an NVIDIA B200, at an inference cost below $0.01, because its feed-forward design produces the 3D representation directly with no per-scene optimization. It accepts up to 64 input views as a spatial context window, so each newly expanded view informs the next.
Viggle also pitches JST 2 as a way to model real places. The company says it reconstructs complete 3D environments from sparse real-world observations, preserving both structure and appearance, and offers that as a way to build and expand robot simulation environments from real locations, with potential applications in autonomous driving, mapping and construction. Scene collisions and object interactions still rely on rule-based physics. Viggle is developing JST-3 to replace those rules with learned dynamics, toward what it calls an experience engine for physical AI, where agents learn by acting in simulated worlds and observing the consequences.
Characters carry their motion in the same tokens. Viggle says the dynamics vectors encode how each part deforms, so characters can be animated without a separate rigging and skinning step, and it reports blind human evaluations in which JST 2 led Meshy 7.1 and Tripo H3.1 on character generation, and Cartwheel, HY-Motion-1.0 and DeepMotion on motion from text or video.
JST 2 runs inside PINOC, Viggle's web app, through Agent Mode, hand editors or an outside agent over MCP. Agent Mode takes a text description and generates the characters, motion and scene, or takes a video and extracts the scene and the moving character from it. Separate Scene, Character and Motion editors handle each piece by hand.
The Scene editor builds an environment from text, an image or a panorama, then places a character in it that can perform any motion and be walked with WASD. An Expand button fills in blurry or black patches so the scene holds together from every angle, and a Geometry-Guided Panorama skill turns a single scene image into a panorama to use as the source. The Character editor turns one line of text or a photo into a rigged 3D character, and the Motion editor produces animation from text or any video. Recently, I have tried creating both my own character and generating scenes from normal 2D photographs.
Claude Code, Codex, Cursor and other agents can connect to PINOC through the PINOC MCP server, which signs in through the user's own MCP client and needs no API key. Viggle lists a character at 10 credits, or 6 with refinement off, motion from text or video at 1 credit per second, and a scene at 5 to 11 credits, with preset library clips free. Motions download as GLB, FBX or USDZ for Unity, Blender and other tools, while characters, which are stored as .vsplat splats, render only through Viggle's free Splat Engine plugin for Unity or the web.
The public JST-2 gallery lets you walk through its scenes in 3D. Like World Labs’ Marble, JST-2 outputs its worlds as Gaussian splats, but it generates a scene in about a second, where Marble takes minutes, and the same model also generates the characters and motion that move through them.
JST 2 is available now in PINOC, with a free plan that includes 60 credits.
