VAST, the company behind Tripo, has open-sourced EdenLive, a real-time generative renderer for interactive worlds. The 5B and 14B weights are on Hugging Face, and the inference code and a technical report are public too. VAST says the 5B model can generate 720p video at 24 frames per second or more on a single RTX 5090 and keep going for minutes at a time.

EdenLive doesn't simulate the world on its own. A game engine or a coding agent runs the rules, tracks the state of the world, and outputs a rough reference video. EdenLive takes that video, a target first frame and a text prompt, and turns it into detailed frames as it plays. Movement, collisions, health bars and win or lose feedback all stay with the upstream program. VAST says end-to-end latency, from a key press to the matching frame showing up, is currently in the hundreds of milliseconds.

According to the technical report, the video generator starts from Wan2.2-TI2V-5B and Wan2.1-I2V-14B models. Along with the weights, VAST released streaming inference code, a skill that turns a text prompt into a rough playable game, and a web front end for the whole prompt-to-game pipeline. That front end calls the Claude CLI to build the games.

The code and the weights don't share a license. The GitHub code is under Apache 2.0, while the weights on Hugging Face use a mixed license called edenlive-mixed, with different terms for different parts. One part that only the web front end uses is limited to non-commercial use. Anyone planning commercial use should read the LICENSE files first.

On hardware, VAST lists Linux with 80 GB NVIDIA GPUs as the tested setup for its original code, and the full set of weights is about 210 GB. VAST's own README lists the single RTX 5090 result, which uses inference optimized with LightX2V, a separate acceleration project. LightX2V's guide reports about 24.7 FPS for the 5B model at 1280×704 with INT8 quantization, peaking at about 28.4 GiB of GPU memory.

VAST also names what still needs work. The report says the model only infers which object in the reference stands for which character or item, that its short memory can lose fine details after something is hidden for a long time or a scene is revisited, and that users can't yet adjust how closely it follows the reference.