Viggle JST-2: 3D characters that need no rig, and motion your agent can ask for

Viggle launched JST-2 this morning (Oct 19, about 9:20 AM MDT) and it was the biggest AI-video launch in our feeds today. The launch post had 885 likes and about 643K impressions when we checked, and a wave of write-ups followed within hours. Most world models learn from video pixels. JST ('Joint Space-and-Time') stores a scene as persistent 3D 'atoms' that keep their identity as they move. They're decoded into Gaussian splats that render locally in your browser. Viggle says the first scene takes about a second and costs under a cent of inference on an NVIDIA B200, counting model inference only.

Ignore the world-model framing for a minute. The character and motion parts are what matter for animation and AI-film pipelines. From one image or a text prompt, JST-2 builds a full-body 3D character whose deformation is learned, so there's no separate rigging or skinning step. Viggle says it also fills in bodies from cropped or partly hidden photos. Motion comes from text ('one sentence returns a motion') or from ordinary video, and the hidden parts of the body are inferred. Viggle says those motions apply to JST-2 characters without retargeting and also work with standard 3D rigs. The PINOC app exports motion as FBX or GLB for Blender, Unity, Unreal or Maya.

The agent angle is why it's on ReelStack. PINOC has a remote MCP server that you connect with a Google sign-in. Per Viggle's own docs it has ten tools, including list_motion_library, generate_motion, generate_character and wait_for_tasks. Generated motion costs 1 credit per second and returns 4 samples, and a character costs 6 credits (10 with refine). There's also a REST API: text-to-motion is $0.10 a request, motion from video is $0.05 per source second, and a rigged splat character is $0.25. That points to a code-first workflow: your agent writes the shot list, pulls a motion for each beat, and drops the result into Blender or a web scene. Viggle's @viggle/splat-engine npm package loads .vsplat characters into web apps. We haven't tested whether it works inside a Remotion or HyperFrames render.

Read the benchmarks as vendor numbers. In Viggle's blind tests, JST-2 beat Meshy 7.1 on character generation with a 58.4% average score, and beat Tripo H3.1 with 73.2%. On animation readiness its scores were 73.8% and 88.9%, with hands the clearest win. Viggle also lists the limits plainly: holes or merging at shoulders, armpits and fingers, detail that softens in close-ups, trouble with unusual body proportions, and clothing and hair that don't swing or collide. Scenes are static, and learned physics is promised for JST-3. Their rules forbid deceptive impersonation, so use your own characters or consenting people. Spotted, not tested. Credit Viggle Research.