VAST Open-Sources EdenLive, a Real-Time Research World Model

Picture this: you walk up to a door, unsure whether you can get through, so you turn around, take two steps, then jump twice to see whether you can clear the wall beside it. In games, this is about as ordinary as it gets. But if the picture in front of you is generated by AI in real time, the model has to keep up with every spur-of-the-moment decision the player makes (suddenly stopping, suddenly turning, suddenly running and jumping), and every frame must remain stable. This is the challenge that VAST's newly open-sourced real-time world model EdenLive sets out to solve.

The division of labour is the key. The game's rules, health, collisions and win conditions are still maintained by the upstream interaction system: when the player presses the attack button, it first determines whether the kick connected and how much health should be deducted, then updates the character's state and renders each change as a grey-box image containing only spatial layout and motion. EdenLive takes over this grey-box image and lets the generative model decide what it ultimately “looks like”: a reference image determines structure, motion and camera, a target first frame fixes materials, colour and style, and text prompts add detail. In Project Eden, which VAST published in May this year, the company laid out this three-layer architecture as a complete scheme.

The team used Gemini-3.8-Flash to run pairwise blind evaluations on 45 long-form cases. Across five dimensions (reference adherence, first-frame consistency, physical plausibility, visual quality and avoidance of appearance leakage), EdenLive 14B and 5B ranked first and second across the board. Take reference adherence: the 14B scored 1,704, about 200 points ahead of AgentGarten. On speed, the team reports that EdenLive 5B can generate 720P video at over 24 FPS on a single consumer-grade RTX 5090 graphics card, and supports continuous generation lasting minutes. These figures come from the team's own evaluations and disclosures, and independent verification is still some way off.

For game prototyping and animation preview teams, the change is concrete. With the same grey-box scene, swapping in a new first frame and a single sentence of prompt lets you compare completely different looks: realistic, hand-drawn line art, cartoon, pixel art or oil painting. When the camera leaves and returns, the scene's structure is still in place. For robotics teams, a video of someone wiping a table, shot with a head-mounted camera, can have the human hands replaced by robotic hands while keeping the motion and camera work, supplementing training data from the “robot's-eye view.” Model weights, inference code and the front-end workbench are now open, though the source material gives no timeline for large-scale deployment.

Its limitations are equally clear. EdenLive does not judge what happens in the game, nor does it make the world run on its own. An existing recording that has been re-rendered is still just a video, and playability comes from the upstream code. Past world-model demos often used a video to prove that a model can imagine a world. EdenLive's novelty lies in separating “who manages the rules, who manages the picture,” and letting the picture move along with the player's inputs.