VAST, the 3D generative AI company behind the Tripo platform, has open-sourced EdenLive, a real-time model that takes simplified game footage or an existing video and continuously re-renders it as detailed frames 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.
EdenLive does not simulate the world on its own. An upstream program, such as a game engine or code written by a coding agent, runs the rules, tracks the world state and outputs a coarse reference video. EdenLive then renders that reference, guided by a target first frame and a text prompt. VAST argues that this split keeps game logic explicit and testable, and lets the limited capacity of a real-time model focus on appearance. It first set out the approach in June as part of Project Eden, its longer-running world model research effort.
According to VAST, the 5B model runs above 24 frames per second at 720p on a single Nvidia RTX 5090. The GitHub README puts the figure at about 24.7 FPS at 1280 by 704 pixels, using acceleration from the LightX2V inference framework. The technical report lists 29.8 FPS at the same resolution on one H100, and 26.8 FPS at 832 by 480 for the 14B model across eight H100s. VAST says end-to-end latency from a user's input to the matching frame is in the hundreds of milliseconds, and that the system can generate continuously for minutes. The two models build on the Wan2.2-TI2V-5B and Wan2.1-I2V-14B video models.
The inference code and the core diffusion, VAE and decoder weights are under Apache 2.0, but not every component is. The vision encoder follows the DINOv3 license, and the image model used by the prompt-to-game workbench is under the FLUX Non-Commercial License, so that workbench is for non-commercial use only. VAST also released Game Prototype Turbo, a skill that turns a text prompt into a coarse playable game that EdenLive can then render.
VAST's demos include a platformer in which a character freezes lava to create footholds, a two-player fighting game whose health bars and round timer are tracked upstream, and video re-rendering that turns a person filmed against a blue screen into a cartoon character. It also shows the model replacing human hands with robot hands in first-person footage to create extra training data for embodied AI.
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 all still need work.
SEE ALSO: AIsphere Launches PixVerse R2, a Real-Time World Model That Remembers What Happens in a Session · Kling 4.0 Goes Fully Live on klingai.com With Audio, MCP and CLI Hooks