VAST open-sources EdenLive, a real-time video renderer
The 5B model turns rough game footage into detailed frames at more than 24 frames per second on a single RTX 5090; the 14B build runs at 480p.
Key takeaways
- VAST open-sourced EdenLive, a generative renderer that turns a rough proxy stream from a game engine, 3D scene or recording into detailed video as it plays.
- The release puts 5B and 14B model weights on Hugging Face, with code under the Apache 2.0 license.
- VAST says the 5B model renders 720p at about 24.7 frames per second on a single consumer RTX 5090, and can generate continuously for several minutes.
- Full real-time mode needs Linux, an 80 GB GPU and CUDA 13, and the project ships no cloud API.
VAST, the company behind the Tripo 3D platform, has open-sourced EdenLive, a generative renderer that turns a rough video stream into detailed footage in real time. The release puts 5B and 14B model weights on Hugging Face and the code under the Apache 2.0 license, with a technical report and demos.
The idea is to split a scene's logic from its look. A game engine, a coding agent or a simulation decides how a scene evolves, while EdenLive generates the visuals. Its authors describe three layers: continuously evolving world state, a state-to-visual interface, and a generative renderer.
How it runs
EdenLive takes a proxy stream, meaning frames from a 2D or 3D scene, a game engine or an ordinary recording, plus a target first frame and a text prompt, and re-renders the frames in a chosen style, from photorealistic to anime. It ships two branches: Wan2.1-I2V-14B at 480p and Wan2.2-TI2V-5B at 720p.
The project page claims more than 24 frames per second at 720p on a single consumer GPU. In the authors' own measurement, the 5B model reaches about 24.7 FPS at 1280x704 on an RTX 5090 through the LightX2V accelerator. VAST says EdenLive can also generate continuously for several minutes.
What it costs to run
The real-time service uses a TAEHV decoder, and the 5B model is the practical entry point. The catch is the setup. Full real-time mode requires Linux, an NVIDIA GPU with 80 GB of memory and CUDA 13, well above the "single consumer card" headline. There is no cloud API; EdenLive runs self-hosted. The main code and the diffusion and VAE weights are Apache 2.0, but the FLUX-based first-frame component is distributed under non-commercial terms.
Why it matters for creators
Separating state from rendering opens workflows that pure text-to-video cannot: interactive game previews coupled to a coding agent, scene and animation previews where a creator swaps the look while keeping the layout and camera path, real-time restyling that replaces characters or a whole style mid-shot, and robot-learning work that converts human manipulation footage into a robot's viewpoint. Tripo AI, VAST's product brand, framed the release as a milestone in its Project Eden world model.
Weights are available to download now; VAST also points to hosted demos on fal and Reactor, and the 3D tool Unbound Loop has integrated EdenLive so users can try it there.
Download the weights with hf download VAST-AI/EdenLive; the real-time service instructions are in the repository's realtime/README.md.
Sources
- tripo3d.ai - the maker's project page: layers, resolutions, FPS claim
- huggingface.co - weights, model list, decoder and encoder components
- github.com - the code and real-time service instructions
- pandaily.com - VAST as the company behind the Tripo platform