Viggle Launches JST-2, a World Model That Generates 3D Scenes, Animation-Ready Characters, and Full-Body Motion

Viggle launched JST-2, a physics-native world foundation model that generates explorable 3D scenes, animation-ready characters from a single image, and full-body 3D motion from text or video, according to the company's JST-2 page. The company says all three outputs share one 3D world: characters animate without separate rigging and skinning, and JST-2 motions apply directly to JST-2 characters without a separate retargeting step.

• Scenes arrive as Gaussian splats. A text prompt or 360° panorama returns an explorable .spz splat file that can be viewed from new angles without rerunning the model.
• Characters export as splats, not meshes. One image becomes a 3D character with learned motion dynamics, delivered as a .vsplat asset that renders through Viggle Splat Engine.
• Motion exports to standard rigs. Text and video inputs produce GLB or FBX animation for Mixamo, MetaHuman, or MetaHuman Body skeletons.
• API pricing varies by job type. Viggle lists $0.06 for a scene from a prompt, $0.25 per character, $0.05 per source second of video motion, and $0.10 per text-to-motion request.
• Physics is only partly learned. Character motion uses learned dynamics, while scene collisions and object interactions still rely on rule-based physics.

Viggle claims scene inference takes about a second, and the splats render locally in a browser

According to Viggle, JST-2 generates a 3D scene in about one second at an inference cost below $0.01, using a feed-forward architecture that skips per-scene optimization. That timing covers model inference on NVIDIA B200 hardware and excludes network latency; the company's own chart puts Marble at about five to six minutes.

The sub-cent figure covers inference only. The API rate card charges $0.01 per new view and counts a panorama as six views, so a scene built from a prompt costs $0.06.

• Up to 64 input views. The model treats them as a spatial context window, so each expansion adds context for the next view.
• Local playback. Splats render in the browser and can be explored repeatedly; the model runs again only when new content is generated.
• Real-world reconstruction. Viggle says JST-2 can also reconstruct complete 3D environments from sparse real-world visual input.

Planned updates include text-guided scene expansion and a version with more splats per scene for finer detail.

Characters and motion are built to work together without retargeting

JST-2 models the world as persistent 3D "atoms" encoded as Physics-Native Tokens, each holding an xyz position plus learned vectors for attributes and dynamics. For characters, those dynamics vectors encode how each body part deforms under motion, which the company credits for animation without separate rigging and skinning.

The model can complete characters from cropped or partially occluded images across a wide range of visual styles. For video input, JST-2 reconstructs visible movement and infers how hidden body parts move.

Export formats shape how these assets fit an existing pipeline. The .vsplat character is a Gaussian splat asset, not a mesh for GLB or FBX importers, so Viggle suggests retargeting Mixamo motion to your own mesh or matching joint sets when pairing a generated character with MetaHuman motion.

In Viggle's blind tests, JST-2 scored 58.4% against Meshy 7.1 and 73.2% against Tripo H3.1 on character generation, where 50% means equal preference. For motion, the company reports 74.8% against Cartwheel Swing on text-to-motion and 58.1% against Cartwheel Comic 4 on video-to-motion, with all clips played on the same untextured MetaHuman rig.

Viggle's character benchmark also lists limits. Hair, clothing, and materials soften as the camera moves in, and clothing and hair deform with the body but are not simulated, so they do not swing from inertia or collide.

• Motion headroom. Viggle's motion benchmark says fast, full-body, strongly periodic motion is where the model has the most room to improve.
• Planned updates. Upcoming releases target better motion accuracy, reconstruction of a person's path through a scene from video, and improved clothing and hair movement.

Access runs through PINOC, the API, and MCP-compatible agents

JST-2 is available free in the browser through Viggle's PINOC app, where scene mode generates 3D scenes and agent mode combines scenes, characters, and motion. Developers get 100 free API credits, with one credit equal to $0.01.

Viggle also offers an MCP connection for Codex, Claude Code, and other MCP-compatible agents. Per the company, API inputs and outputs are not used to train its models, outputs can be used commercially when you hold rights to the submitted material, and signed output links expire after one hour.

JST-2 builds on JST-1, which validated the architecture on character generation and animation and powered Viggle's meme-making app, now past 50 million users by the company's count. Viggle is developing JST-3 to replace the remaining rule-based collisions and object interactions with learned dynamics that predict how objects interact.