JST-2, Viggle's World Model, generates a scene in one second

Designed in 3D from end to end, Viggle's JST-2 generates a scene in one second and animatable characters without rigging, but maintains rules for collisions.

According to Viggle, many World Models learn the physical structure of the world indirectly, from the appearance of images. Known for its animated meme app, which claims over 50 million users, the startup took the opposite approach with JST-2, which it presents as a "physics-native" foundational World Model. Four years of work, including two dedicated to an initial version already focused on character animation, culminate in an architecture built entirely in three dimensions, from data encoding to training. Scenes, characters, and movements are generated within a single 3D world that can be navigated and expanded. The model can be used for free in the browser via the PINOC app, whose free tier provides 60 non-renewable credits without commercial use, and developers can also access it via API at $0.01 per credit. Persistent atoms rather than images

The name JST stands for Joint Space-and-Time. The model describes a scene as a set of persistent atoms, each corresponding to a piece of matter or an object part that retains its identity as it moves or deform. Each atom possesses an explicit xyz position and learned vectors describing its attributes and dynamics, all encoded as Physics-Native Tokens. An encoder produces these tokens from real images, a generator creates them directly from multimodal prompts, and a decoder renders them into images or 3D Gaussian splats depending on the viewpoint and the applied movements or forces. A scene in one second, explored without recalculation

According to the launch post, a 3D scene is generated in about one second of inference on an NVIDIA B200 GPU, network latency excluded, for a cost of less than a cent. The architecture directly produces the 3D representation without scene-specific optimization, whereas Viggle measures five to six minutes for Marble, World Labs' tool, without detailing the conditions of that test. Once generated, the scene is decoded into Gaussian splats rendered locally in the browser, and the model is only called upon to create new content. Up to 64 views serve as a spatial context window: each new viewpoint completes the world in real time and enriches the context for the next view.

The same mechanism reconstructs full environments from a few real observations, where point cloud reconstruction only yields coarse geometry according to Viggle. The company sees this as a way to build simulation environments for robots from existing locations, with potential uses in autonomous driving, mapping, or construction. Animated characters without rigging

From text or a single image, even cropped or partially masked, the model generates a 3D character whose tokens already contain how each part deforms in motion. Because this dynamic is learned from data, the character animates without the separate rigging and skinning steps, which consist of attaching a skeleton to it and then binding its surface to it. Movements, generated from text or reconstructed from video, apply directly to these characters without retargeting and remain compatible with standard 3D skeletons. From a video, the model also infers the motion of body parts that are masked or off-screen.

The published comparisons come from blind human evaluations, whose score adds the win rate and half of the ties, with 50% indicating equal preference. In character generation, the model achieves 58.4% against Meshy 7.1 and 73.2% against Tripo H3.1, with its best results on faces, and the gap widens for ease of animation, at 73.8% and 88.9%. For motion, it surpasses Cartwheel Swing (74.8%) and HY-Motion-1.0 (80%) from text, then DeepMotion Animate 3D (89.1%) from video. Against Cartwheel Comic 4, the score drops to 58.1%, with 53.1% ties, and the motion benchmark page acknowledges that this competitor remains ahead on fast sports videos and that feet can still slide during footplants. Neither the number of evaluators nor the sample size are specified. Still rule-based physics for objects

The "physics-native" designation currently has a limited scope. In JST-2, learned dynamics animate characters, but scene collisions and object interactions still rely on rule-based physics. The upcoming JST-3, currently in development, is intended to replace these rules with learned dynamics capable of predicting how objects interact and transform the world. The company aims for an "experience engine" for physical AI, where agents would learn by acting in simulated worlds and observing the consequences of their actions.