VAST Open-Sources EdenLive, a 5B Real-Time World Model That Tops 24 FPS at 720p on a Single RTX 5090

VAST, the Chinese startup behind the Tripo 3D-generation family, has released EdenLive, a real-time world model that turns simple proxy scenes or reference video into continuously rendered, interactive visuals — and it has open-sourced the whole thing. Model weights at multiple sizes are on Hugging Face, inference code is on GitHub, and the company published a technical report alongside an end-to-end demo.

The headline numbers are all vendor-reported: the 5-billion-parameter model generates at more than 24 frames per second at 720p on a single consumer RTX 5090, sustains minute-scale continuous generation, and responds to user input with end-to-end latency in the hundreds of milliseconds. Output ships in 480p and 720p tiers. VAST also claims EdenLive is the first model to combine arbitrary reference inputs, arbitrary art styles and long continuous operation under real-time constraints — a claim that has not been independently verified.

The architecture rests on a division of labor that VAST laid out in its Project Eden research preview in June: game logic and world state live in an upstream interaction system, while the generative model handles only the rendering. Players press keys, the interaction system updates positions, health and rules, and EdenLive receives simplified reference frames describing layout and motion, then paints them as detailed visuals in real time.

Three inputs steer the output: a reference scene or video supplies layout, camera and motion; a target first frame sets the overall look and art style; and a text prompt fills in stylistic detail. Because those channels are separate, the same blocky proxy level can be rendered as a photorealistic scene, an anime sequence or a hand-drawn sketch without touching the underlying game logic.

The demos VAST published show where the company thinks this goes. Interactive game previews let creators roam low-frequency gameplay like puzzle levels with rendered visuals instead of gray boxes. Scene and animation previews swap art styles on fixed camera paths for directors and concept artists. Video re-rendering replaces a live-action presenter with a cartoon character while keeping the motion, or converts real footage of a house into a soft animated style. Most notable for robotics researchers: embodied-data augmentation, where first-person videos of human hands are re-rendered with robotic grippers, materials and objects changed while the manipulation motion stays intact — a cheap way to multiply training footage for imitation learning.

VAST is candid about the limits. High-frequency, high-intensity interactions still outrun the generation speed; identity preservation degrades when several characters share the frame; small objects resist fine control; and visual consistency drifts over long generations. The company is also releasing Game Prototype Turbo, a prompt-to-proxy-game skill that generates a playable prototype whose state stream feeds EdenLive's renderer, closing the loop from natural language to a rendered, playable game.

The release lands in a week when world models are crowding the agenda — Google DeepMind's Genie 3 generates playable 720p worlds at 24 FPS behind closed doors, and Chinese robotics labs are bolting world models onto VLA systems for physical prediction. EdenLive's bet is different: that real-time generative rendering becomes useful only when the weights, the code and the full pipeline are open enough for game developers and embodied-AI teams to plug in their own state systems. At 5B parameters and one GPU per scene, that bet is now cheap enough to test.