ReWorld

Explore, leave, return — the world remembers.

A real-time interactive world model with long-horizon memory

Real-time interactive 704×1280 streaming Long-horizon memory
Method

How it works

One streaming diffusion backbone, trained so that control and memory stop fighting over the same context window.

ReWorld framework diagram: window-split training with head routing, an odometer-admitted landmark memory bank with pose retrieval, and a distilled real-time streaming decoder.
The ReWorld framework — split training windows, routed attention heads, and a bounded landmark memory queried by camera pose.
1

Window-split training

Control learns under short windows, memory under long ones. Per-step head routing decides which attention heads see the distant past, so precise action-following and long-range recall are trained without compromising each other.

2

Bounded memory

An odometer-admitted landmark bank with pose retrieval. Views are admitted as landmarks by traveled distance and retrieved by camera pose under a fixed KV budget — memory stays flat no matter how far you roam.

3

Real-time streaming

Metric-aligned multi-source data, then 4-step distillation. Trajectories from many sources are aligned to a common metric scale, and the distilled model streams interactive 704×1280 video in real time.

Long-horizon memory

The world remembers

Leave and come back — the scene is still there. Twenty-four 16-second journeys: each clip roams away from its opening view, and the badged ones navigate all the way back. Hover any card for its generation prompt; the corner button opens it full-size.

Crystal Depths dolly
Portal Complex strafe
Silver Ruins
Turtle Realm dolly
Aurora Fortress dolly
Robot Factory dolly
Bear Conductor yaw
Frog Explorer dolly
Station Traveler dolly
Cathedral Nave
Coral Reef strafe
Miniature Railway L-return
Noodle Street
Storybook Lakeside
Treasure Caves
Elven City yaw
Stone Castle
Pixel Labyrinth
Cloud Boulevard Walk
Alchemy Tower
Clockwork Park dolly
Fairy Flight dolly
Mars Exploration Rover
Alien Saltflat Hovercraft
Camera control

Follow any camera intent

Twelve clips driven by live key input — the exact sequence is shown in the on-screen HUD — while the scene stays coherent, in first person and third.

Floating Isles
Alpine Snowfield
Watercolor Port
Cloud Boulevard
Submarine City
Balloon Archipelago
Owl Librarian
Desert Rally Truck
Lunar Pressurized Buggy
Solarpunk eVTOL
Floating City Catamaran
Asteroid Ring Tug
Metric scale

Learning the world beneath the pixels

Every training source shares one physical action scale, so the model has to internalize real size and distance, not just appearance: the same key press always covers the same meters. Far drifts, near rushes — from high altitude the world barely crawls, up close it streams past. Each row pairs the two extremes — same keys, same meters, only the distance to the world changes.

Alpine Peaks
Rural Farmhouse
Seaside Town
Toy City
Coastal Flight
World Tree Balcony
Long horizon

Generation Beyond

Minute-scale rollouts from a single starting frame. Bounded memory keeps the stream going — the world keeps unfolding.

01 Desert Astronaut Third person — an astronaut treks the dune field, rover convoy and ridgeline persistent.
02 Alpine Snowfield — Long Weaving shuttle runs over wind-carved snow — ridgelines and rock bands return to view intact.
Hero shot
03 Cloud Teahouse Village A steady glide across wooden bridges over a sea of clouds into a golden-hour floating teahouse village, ending inside a lantern-lit porch.
04 Robot District A bright futuristic plaza with robots, moving walkways and holo-screens, ending in a clean tree-lined courtyard.
05 Enchanted Wood A lush glowing forest path with bioluminescent flowers and a stone bridge — vivid, dense, instantly readable.
06 New Realm An epic vista through a giant ring portal onto a waterfall valley city — the strongest single composition in the pool.