VAST Open-Sources EdenLive Real-Time World Model for Single-GPU 720p Generation

According to VAST, EdenLive can take a simplified scene or reference video, a target first frame and a prompt, then follow the reference layout and motion to generate detailed images in styles such as photorealism or anime. Reference sources can differ and the output style can be specified, with those changes shown during real-time generation. On a single Nvidia RTX 5090 at 720p, the 5B model generates more than 24 frames per second and can run for minute-long sequences. The end-to-end delay from user input to the corresponding image is at the hundreds-of-milliseconds level, VAST says.

The model is positioned as a rendering layer for interactive content. An upstream interaction system receives player actions and updates world state; EdenLive uses the corresponding reference images to present those changes. VAST says it has released model weights and inference code, along with a technical report and end-to-end demos. The code is available on GitHub, weights on Hugging Face, and project information on VAST's Eden site.

In the demos described by QbitAI, EdenLive is used for interactive game previews in 2D and 3D scenes, including a lava course in which a character uses a freeze gun to make footholds and a two-player fighting game where character actions are paired with health bars and round timers. VAST says the current system is better suited to lower-frequency interactions such as scene roaming and box-pushing prototypes, where creators can operate a game prototype and see a detailed visual version.

EdenLive also supports scene and animation previews. With layout and camera movement fixed, the same input can be rendered in different art styles; the demo includes a pelican riding a bicycle and a snow-driving scene built from simple models, with generated textures, snow and lighting. For video re-rendering, existing footage can be converted in real time to replace a character's appearance or the overall style while keeping posture and clothing colors consistent with the reference. A live-action house can be turned into hand-drawn animation while retaining its layout.

For embodied data augmentation, EdenLive can turn real human operation footage into a robot viewpoint and alter objects and appearances in first-person video. In one demo, hands holding a spray bottle and cloth are replaced by robotic hands and the cloth changes from orange to yellow while the cleaning motion stays the same. In another, human hands become a metal robotic arm and a cart handle takes on a silver metallic finish, with grasping and pushing motions preserved. VAST says these uses can expand video samples for embodied research.

The technical route is to decouple world state from visual rendering. The interaction system maintains state and rules, while EdenLive handles visual generation. VAST's earlier Project Eden research preview, published in early June, already made this decoupling a core design; VAST says the choice came about three to four months before an industry consensus formed. Project Eden uses a three-layer architecture: continuously evolving structured state, an interface from state to observation, and generative neural rendering.

VAST says its data pipeline covers paired data from 2D and 3D simplified scenes, engine renders and real video, aligning scene layout, camera motion and object positions. The team spans 3D, video generation and system engineering. EdenLive still has limits: high-frequency, intense interactions demand more from generation speed and motion; identity preservation with multiple characters, fine control of small objects and visual consistency over long generations also need further work.

VAST also open-sourced Game Prototype Turbo, a Prompt to Proxy Game Skill that lets users quickly generate playable game prototypes. When the game runs, the interaction system updates state and outputs simplified reference frames, and EdenLive renders them into detailed visuals in real time, forming a link from game logic to real-time visual output. VAST says its longer-term Project Eden effort aims to let people freely create and experience interactive worlds, with EdenLive advancing the real-time generative rendering part.

Editor's Summary

VAST has open-sourced EdenLive, a 5B real-time world model that renders 720p interactive scenes above 24 FPS on a single RTX 5090 and supports arbitrary reference inputs, selectable styles and minute-long generation. Its decoupled design separates state maintenance from visual rendering, while VAST acknowledges limits in high-frequency interaction, multi-character identity preservation and long-run consistency.