VAST open-sources EdenLive, a real-time world model rendering game frames at 24FPS on an RTX 5090

APPSO · wechat · 2026-10-20

VAST has open-sourced EdenLive, the real-time generative renderer from its ProjectEden world-model stack, releasing 5B/14B weights, streaming inference code, a tech report, and EdenGolem v0.1.

The key design is a division of labor: the upstream interaction system maintains game state, rules, collisions, and HP, and renders each update as a gray-box frame; EdenLive then generates the final visuals in real time from that plus an optional first frame and text prompt. References can be a rough 3D whitebox, 2D scenes, or real footage—swapping the first frame or prompt yields realistic, cartoon, pixel-art, or painted looks. The team claims the 5B model sustains 720P at over 24 FPS on a single RTX 5090 with minute-scale continuous generation, and blind Elo pairwise evals (45 long-horizon cases, judged by Gemini-3.8-Flash) put EdenLive 14B/5B first and second across five dimensions including reference adherence.

The write-up also highlights embodied data augmentation—rewriting egocentric human videos into robot-arm viewpoint footage for robot training—and EdenGolem, a skill that lets coding agents build runnable 3D gray-box games from a gameplay description. A V·STAR talent program for 3D and world-model research launched alongside.

Original post →

More from Multimodal

Multimodal channel →