VAST Open-Sources EdenLive Real-Time World Model

EdenLive takes three inputs. A reference video determines scene layout, camera movement, object trajectories and event timing; a target first frame sets the subject, materials, color palette and visual style; and a text prompt adds style and detail, with additional descriptions possible during interaction. The reference can be a rough 3D white model, a 2D scene or existing footage. Changing the first frame and prompt can produce realistic, line-drawing, cartoon, pixel or oil-painting looks while keeping the same structure and motion.

The release includes 5B and 14B model weights, each in bidirectional and few-step streaming versions, real-time streaming inference code, EdenGolem v0.1 and a front-end workbench, plus a technical report. VAST said the 5B model can generate 720p images at more than 24 frames per second on a single consumer-grade RTX 5090 graphics card and support continuous generation for minutes.

In a blind evaluation using Gemini-3.8-Flash on 45 long-duration cases, with pairwise comparisons and order swapped, EdenLive 14B and 5B ranked first and second across five dimensions: reference adherence, first-frame consistency, physical plausibility, image quality and avoidance of appearance leakage, according to the report. On reference adherence, the 14B model scored 1,704, about 200 points above AgentGarten.

EdenLive does not maintain game rules or decide what happens in a scene. In Project Eden, which VAST released in May, the world model is divided into three layers: a continuously evolving world state that records object positions, attributes, interaction rules and action outcomes; an interface that converts state into layout, geometry and event constraints for a given viewpoint; and a generative renderer that produces high-fidelity images. EdenLive implements the latter two layers and leaves the first to an upstream interaction system.

In a fighting game, for example, a player's attack command still goes to the upstream system, which determines whether the kick lands, how much health is lost and where the characters move. That system then updates the state and renders a new gray-box frame. A fine-tuned visual encoder extracts layout, camera and motion from the gray-box frame, and EdenLive renders the final visuals. VAST says this division keeps deterministic rules out of the video model while allowing visual elements to remain stable as the camera moves away and returns.

The model can be used for game rendering, animation previews, style transfer and embodied data augmentation. In robot training, VAST says, first-person footage of a person wiping a table or pouring water can be converted into a robotic-arm view while preserving the action, object positions and camera movement, replacing human hands with a mechanical gripper. For game development, VAST also released EdenGolem v0.1, a skill for coding agents. Given a gameplay description, it completes gameplay design, scene construction, interaction logic and interface work in four stages and produces a standalone 3D gray-box game. It supports puzzle, shooting, driving, adventure, arena and fighting games.

The report notes limits. Existing footage can provide camera and motion, but after rerendering it remains a video, and playability comes from upstream code. VAST's separation of state and rendering still needs testing over longer and more complex interactions. The report also cited a January 2026 Oxford review that listed "visual confusion" as a challenge for world models, arguing that realistic images do not by themselves show that a model understands causality or physics. VAST also announced a V-STAR talent program for 3D generation, 3D understanding and agentic AI, and world models.