VAST open-sources EdenLive, a real-time research video model

A character freezes a patch of lava to make a foothold, and the same jump, filmed as a rough grey sketch, comes back as a detailed scene. That is the kind of demonstration VAST put in front of developers this week. The company, behind the Tripo platform, has open-sourced EdenLive, a real-time model that takes simplified game footage or an existing video and re-renders it frame by frame in a chosen style, from photorealistic to anime. The release includes 5B and 14B model weights, real-time streaming inference code, a technical report and end-to-end demos.

The model works as a second stage. An upstream program, such as a game engine or code written by a coding agent, runs the rules and tracks what is happening in the game. It then outputs a coarse reference video. EdenLive takes that reference, starts from a target first frame and a text prompt, and paints the detail on top. VAST says this split keeps the game logic explicit and testable, and lets the model's limited capacity focus on how things look. The approach was first presented in June, as part of Project Eden, the company's longer-running research effort on world models.

The speed figures are the striking part. VAST reports that the 5B model runs above 24 frames per second at 720p on a single Nvidia RTX 5090 graphics card. The GitHub page gives about 24.7 FPS at 1280×704 pixels, with the LightX2V acceleration framework. The technical report lists 29.8 FPS on one H100 chip, the data-centre processor Nvidia sells for this kind of work, and 26.8 FPS at 832×480 for the 14B model running across eight H100s. Those numbers come from VAST itself, and no independent benchmark has been published yet. The two models are built on existing open video models, Wan2.2-TI2V-5B and Wan2.1-I2V-14B.

For a developer, the immediate gain is a shorter path from a rough prototype to something that looks finished. A studio or a solo maker could test a scene's look before building its art, and researchers get weights they can modify. The limits are just as concrete. VAST says EdenLive currently suits low-frequency interaction, such as scene walkthroughs and box-pushing puzzles. Fast, intense gameplay, keeping several characters consistent, fine control of small objects and visual consistency over long runs still need work. The licences also draw a line: the vision encoder follows the DINOv3 licence, and the image model behind the prompt-to-game workbench is under the FLUX Non-Commercial License, so that workbench is for non-commercial use only. Commercial teams should check each component before building on it.