Today we are releasing JST-2, the world's first full-stack physics-native world foundation model. Many world models learn in pixel/video space. We went back to first principles and built the entire model stack in 3D. Generate a 3D scene in seconds and expand it in real time. Create characters with... motion built in, all through one general architecture.show more

JST models the world as persistent atoms in 3D space. Each Physics-Native Token has an explicit xyz position and learned vectors for attributes and dynamics. We spent four years building the full physics-native stack around this representation: tokenization, architecture, the data synthesis engine, pre-training, and post-training.

Generate a 3D scene in about one second of model inference, at a cost below $0.01. Move to a new viewpoint and generate the missing parts in real time. Keep expanding the same world, with up to 64 input views as a spatial context window.

From one image to a 3D character with motion built in. Its dynamics are learned from data, without separate rigging and skinning steps. JST-2 outperforms state-of-the-art models in blind human evaluations of character generation and animation readiness. Full benchmark:

Generate full-body 3D motion from text or video. From video, JST-2 reconstructs visible movement and infers how unseen body parts move. JST-2 leads state-of-the-art models on text-to-motion and video-to-motion in blind human evaluations. Full benchmark:

JST-2 reconstructs complete 3D environments from sparse real-world observations and expands them from new viewpoints. This opens a path to building robot simulation environments from real places. Generated scenes render locally, without running the model for every frame.

Next: JST-3. In JST-2, character motion is driven by learned dynamics. With JST-3, we aim to extend learned dynamics to interactions between objects, replacing rule-based simulation with learned physics. Our goal is an experience engine for physical AI, where agents learn by acting in simulated worlds and observing the consequences.

JST-2 is live. Read the research and benchmarks: Try JST-2 for free today.

I had way too much fun building this Lord of the Rings-inspired Rivendell game with ChatGPT Astra + our Viggle's JST-2 world model. Generating the scenes made it so much easier to build, and so much prettier. Come explore the valley, the Unseen World and the Council!

Wow the quality is fire! Crisp scenes, and the characters move so naturally. Let's play!

So proud of the team! JST-1 powered our first product Viggle, which went viral 2 years ago and now has 50M+ users. I can see JST-2 opens up so many opportunities for game development & robotics training. Can't wait for the next milestone!

Can't get to the next milestone without you Nan!

We've spent four years building JST, and I'm excited to finally share JST-2 with everyone. Please try it and tell us what you think! Happy to answer questions about the research or the model. We'd also love to hear what you want to build and explore how we can help!

Four years of conviction and building from first principles brought JST-2 here. Thank you for leading the team through it, Hang. Now we can't wait to see what builders create.

If JST-2 can't spill a cup yet, what are you hoping people build with it while JST-3 is coming?

JST-2 is built for creating usable 3D characters, motion, and scenes today, especially for games, prototypes, and interactive experiences.

Congrats on the launch!

Thank you! We love Drama 3 too 💚

Every JST-2 capability is live in PINOC, free to try 👀 Turn one image into a 3D character, make it move from text or video, and build the world around it. Show us what you make:

Which painting should JST-2 do next? I'd vote Las Meninas

Las Meninas would be an incredible stress test 👀 So many characters, sightlines, mirrors, and layers of depth !

Had a blast generating loads of environments, the details have improved a ton! Look forwards to making more with it.

Yes Roy💪 let's build some killing games!

Turning one image into an animated 3D character without manual rigging could save creators a lot of time.

Exactly. We want creators to spend less time fighting the pipeline and more time building the character and world they imagined.

Is the robotics demo on the JST-2 page somewhere? want to send it to a friend who trains robots

yes, Modeling the real world - this part!

Will give it a read later, looks good

准备试试把JST-2加入到我的游戏制作工作流中！

Looking forward to it!

Had a blast creating scenes with small games using Jst-2 from scene creation to character creation and custom animations with their motion feature built a small game as well!

that's amazing, what game did yo build?

a Small little game playable on browser where you have to pick up all the mushrooms!

beautiful style!

built in 3D first, so when you orbit the room doesn't smear the way video world models do. that's the test I run on every one of these

I’d point an agent at JST-2 and let it keep regenerating the scene till it passes its own check

fair to flag these are Viggle's own blind tests. one model beating four specialists is basically what we saw happen with LLMs though, so I'm not that surprised

one general model that can actually scale. that's our path!

filling in the arm the camera missed is the part I'd test first

A world model predicts how a scene changes from one moment to the next, so you can simulate things before trying them for real

Congrats on JST-2, this is actually sick. 1-second scenes + characters that already come with motion is insane. Gonna go try it right now.😍😍😍

for two person teams rigging ate the whole budget, so most just skipped it. if it's really one image to a rigged character i mostly want to see how the hands hold up

Hands are one of the areas where JST-2 performed especially well in blind evaluations. Definitely put it to the test—and you can find the detailed comparison in our launch blog.

no retargeting step would save me hours on every project

@kirbyman01 sticks and stones may break my bones

This is so good! Great to see Viggle back with amazing model

这个模型的效果太惊艳了~回头试试~

thanks! the 3d character building should be much easier now as it skips the rigging & retargeting

What's the quickest way from a JST-2 scene into Godot or Unity?

This is a massive leap forward. Incredible work! Huge congrats on the JST-2 launch! 🔥

Nice!

Thanks Warper! Welcome to try;)

We’re already putting JST-2 to work in our games. 🎮 Bringing characters to life with high-quality models and motions. See it in action and try our games here:

Super coooooool

Super cool！Big congratulations to the team！

how's JST-2 holding up when it's buried in a long agent loop?

AI search summaries keep citing sources that do not exist

so JST-2 makes the characters inside the same splat scene? Marble leaves me with an empty room to fill

yes exactly, a world with moving characters