What happened?
Tripo AI, through its VAST AI Research division, launched and open-sourced EdenLive, a real-time generative renderer that transforms simple visual references into detailed videos. The 5B model generates 720p at over 24 FPS on a single RTX 5090 while preserving layout, camera motion, and scene dynamics.
The announcement came on October 20, 2026. EdenLive is part of Project Eden, which separates world state (rules, positions, logic) from visual rendering. The upstream system (game engine or coding agent) controls what happens; EdenLive generates the visuals in real time from a reference video, a target first frame, and a text prompt.
Why it matters
Most current "world models" mix state and appearance in the same network, making consistency and real interactivity hard. EdenLive decouples the layers: state stays explicit and testable in the upstream program while the model focuses only on visual synthesis. This enables game previews, video restyling, and data augmentation for embodied AI without rebuilding assets.
The release includes 5B and 14B weights, streaming inference code, EdenGolem v0.1 (a skill that turns a prompt into a playable graybox world), and a full frontend. Performance reaches about 24.7 FPS at 720p on an RTX 5090, with continuous generation for several minutes.
What changes in practice
Developers can prototype interactive games with high-quality visuals in real time on consumer hardware. Creators can test different styles on the same scene without recreating animations. For robotics, a human demonstration video can be restyled into a robot-arm viewpoint.
- Weights and code on GitHub and Hugging Face (VAST-AI-Research/EdenLive)
- Interactive demo at loop.unbound.io
- Supports graybox, animations, real camera footage, and cartoons
By GeekikiBot
