Introducing Project Eden: A World Model for Multiplayer and Agents

We're releasing a research preview of Project Eden, a persistent world model built for multiplayer use and AI agents sharing one consistent world. It runs continuously, remembers what you do to it, and stays consistent no matter which camera looks at it. The full technical preview is here: Project Eden research preview.

The idea behind Project Eden is simple to state and hard to build. Most systems being called "world models" right now are video generators; they predict the next frame and forget everything once it scrolls off. Project Eden keeps the world running underneath the picture instead. Put out a fire and it stays out. Look away from a wall and it's still there when you turn back. Two players race on one track from different angles and stay inside the same reality.

That difference is the whole point. The sections below explain why it matters and how it works.

Why Most "World Models" Aren't

Generative video has gotten good. Action-conditioned models respond to your input and produce smooth motion, and the industry started calling them world models.

There's a gap between predicting pixels and simulating a world, though. Generating the next frame tells you how an image should change. A real world model has to track what those pixels mean: the objects, spaces, rules, memory, and physical consequences that should stay put whether or not a camera is looking.

Research so far has split into two roads, and each hits a wall.

Action-conditioned video generation handles motion well but has no lasting memory. The world's "state" lives inside a short window of recent frames, so anything that leaves the camera can drift, vanish, or come back wrong. You get time without permanence.

Static 3D scene generation gives you solid, walkable space but freezes it. Physics, events, and change aren't part of the design. You get structure without change.

So one road remembers how things move but forgets they exist. The other remembers they exist but can't let them change. A real world model needs both.

The Core Idea: State Before Rendering

Project Eden's main design choice is to split the world state from the rendering.

Reality already works this way. The world is there before any camera looks at it. These are questions about state, not about rendering. So rather than stuffing space, objects, events, and appearance into a stream of pixels, Project Eden keeps an underlying world running on its own. Rendering becomes a way to look at that world, not the place the world is stored.

How It Works: Three Layers

One video generator doing everything gets replaced by three layers, each with a clear job.

An evolving structured state. This is where the world lives. It's a compact representation, not a giant 4D point cloud, and it tracks coarse geometry, object identity, semantics, and what your actions did. Objects that leave the frame don't get tossed. Changes get written back and stay.

A state-to-observation interface. When a view needs rendering, this layer turns the world state into camera-conditioned cues: local geometry, semantics, recent changes. Every viewpoint pulls from the same source, so different cameras stay physically consistent.

Generative neural rendering. The renderer takes those cues and produces what you see: lighting, texture, materials, smoke, fire, water, motion. It doesn't have to remember the world. It only has to show it.

What Eden Can Do

Splitting state from rendering opens up things that pure video or static 3D can't do together.

Objects don't vanish when they leave the frame. They stay in the underlying state, so you can look away as long as you want and the world is still there when you turn back. In our fire-extinguishing demo, putting out the fire isn't a passing effect. The environment enters a changed state and keeps it.

Physics responds to varied input. Steer a boat or drive a car, and the action registers in the state and updates the dynamics.

Worlds are reusable and editable. Video generation runs one way; once the timeline moves on, you can't reach back in. Eden lets you act on a running world over and over. Leave a mark, move an object, trigger a result, and it persists. Other people who enter that world see the same changes.

Multiplayer is built in. Many agents share one compact state, with a separate view rendered for each camera. In the racing demo, two cars share one synchronized track from different angles. In the shooting-range demo, different players take different actions in one environment, and Eden resolves them under the same rules.

It can train agents. A world with stable physics, temporal consistency, and long memory works as an environment for training and testing embodied AI, where actions have reliable outcomes and the world doesn't reset after every glance.

Why It Matters

Eden serves two audiences. For creators, it's an engine for interactive content: generate an environment, set up interactions, invite people into the same persistent space. For researchers, it's a simulation base with long-horizon consistency, real physical rules, editable scenarios, and measurable consequences, which is what embodied agents need to learn reliably.

That's the reason we don't file world models under video generation. A world model needs a state that can change.

What's Next

This is a research preview, not a finished general-purpose world model, and the work is early. We're building richer physics, larger environments, wider free-viewpoint exploration, finer object interaction, and a stronger State Transition Model that updates the world from actions, rules, and feedback. Evaluation has to grow too, testing persistence, causal consistency, rule-following, and multi-agent sync, not visual quality alone.

Moving from predicting the next pixel to simulating the next state isn't only an engineering change. It points toward AI that can create, remember, and reason inside worlds that hold.

Read the full write-up, with demos and architecture detail, in the Project Eden research preview.

About VAST AI Research: VAST AI Research builds 3D foundation models and world models. Learn more at tripo3d.ai/research and follow @vastairesearch.