VAST Open-Sources EdenLive for Real-Time World-Model Rendering

Introduction

Generative models are moving beyond one-off images and short clips toward interactive worlds that respond continuously to user input. VAST’s newly released and open-sourced EdenLive targets a critical part of that transition: turning simplified scenes and changing interaction states into coherent, detailed visuals in real time.

Key points

• Real-time output: VAST reports that its 5B model exceeds 24 FPS at 720p on a single consumer Nvidia RTX 5090. It is also designed for minute-scale continuous generation, with end-to-end input-to-display latency in the hundreds of milliseconds.
• Multiple forms of guidance: EdenLive can use simplified scenes, reference videos, a target first frame, and text prompts. The reference supplies layout, camera movement, and motion, while the first frame and prompt define appearance and style.
• A split between state and rendering: An upstream interaction system maintains character positions, actions, hit detection, health, and other rules. EdenLive turns the resulting state changes into richer visual frames.
• Several use cases: VAST demonstrates interactive game previews, scene and animation previsualization, video restyling, and the conversion of human-operation footage into robot-view data.

Why the architecture matters

EdenLive is not simply a faster video generator. Its more important contribution is placing a generative renderer inside a controllable interaction pipeline. Video models can create compelling visuals from prompts, but they are not naturally suited to reliably maintaining game rules. Conventional engines provide that control, but often require extensive asset production. By separating the two jobs, a logic system can advance a world through blocks, proxy geometry, or other simplified representations, while the generative model upgrades the appearance.

For creators, this could enable a workflow in which gameplay is tested first and characters, materials, and art directions are explored later. The approach may be useful for game prototypes, animation previs, and virtual characters, where repeated modeling and rendering can be expensive. VAST also plans to open-source Game Prototype Turbo, a tool intended to generate playable prototypes from prompts and connect them to EdenLive for visual output.

The limitations are equally important. The current system appears better suited to lower-frequency interactions such as exploration and puzzle-like movement than to fast action gameplay. High-speed motion, identity consistency across multiple characters, fine control of small objects, and visual stability over long runs remain open challenges. In practice, the main engineering problem is not just model quality, but the coordination of state systems, paired data, inference speed, and rendering fidelity.

By releasing weights, inference code, a technical report, and end-to-end demos, VAST gives researchers and developers a concrete system to test. EdenLive does not prove that generative games are solved, but it presents a clear engineering direction for world models: structured systems determine how the world behaves, while generative models determine how it looks.

Source: QbitAI