EdenLive renders 720p video at 24.7 FPS on one RTX 5090
The open-weight renderer's code and main weights are Apache 2.0, but the streaming service needs an 80 GB NVIDIA GPU and CUDA 13.
Key takeaways
- VAST released EdenLive, a real-time generative renderer with open code and weights that turns a rough proxy stream from a game engine, a 2D or 3D scene, or a recording into finished video.
- The 5-billion-parameter model reaches about 24.7 FPS inference throughput at 1280x704 on a single RTX 5090 through the LightX2V accelerator, the authors' own measurement.
- The code and the DiT, VAE, UMT5 and decoder weights are Apache 2.0, while the FLUX-based first-frame component carries a non-commercial license and the vision encoder uses Meta's DINOv3 License.
- The real-time service requires an NVIDIA GPU with 80 GB of VRAM or more and CUDA 13.x on Linux; for readers who would rather not self-host, VAST points to demos on fal and Reactor and an Unbound Loop integration.
VAST released EdenLive, a real-time generative renderer that turns a rough proxy stream into finished video, with the code and the model weights published openly. The company shipped it as part of Project Eden, its research line on interactive world models.
EdenLive takes a reference video, a target first frame and a text prompt. The reference can be a simplified rendering from a coding agent or a game engine, or a recorded clip. The renderer follows the layout, camera motion and timing of that reference and paints it into detailed, stylized video. The design keeps the game logic and world state on the code or engine side, where they can be inspected and edited directly, and leaves the network to render only the visuals.
Control reaches the model through a vision encoder. A post-trained DINOv3 encoder embeds each proxy frame, and those features condition a diffusion transformer. There are two model sizes. A 14-billion-parameter branch built on Wan2.1-I2V runs at 480p, and a 5-billion-parameter branch built on Wan2.2-TI2V runs at 720p. Each comes in a bidirectional version and a few-step, chunk-autoregressive version that samples four steps per chunk for streaming, and the real-time path uses a TAEHV decoder.
On speed, VAST's README says the 5B model reaches about 24.7 FPS inference throughput at 1280x704 on a single RTX 5090 through the LightX2V accelerator. The tech report puts the optimized 5B renderer above 24 FPS at the same resolution on a single H100.
The weights total about 210 GB. EdenLive has no cloud API of VAST's own, but the company points readers who would rather not deploy locally to hosted demos on fal and Reactor, and the 3D asset platform Unbound Loop has integrated EdenLive so users can try it directly. The main code and the DiT, VAE, UMT5 and decoder weights are Apache 2.0. The vision encoder uses Meta's DINOv3 License, and a first-frame component based on FLUX.2 [klein] 9B carries the FLUX Non-Commercial License, loaded only by the real-time workbench.
The catch is hardware. The streaming service's own requirements list an NVIDIA GPU with 80 GB of VRAM or more, CUDA 13.x, Linux and about 100 GB of free storage. That is a far higher bar than the "720p on one consumer card" line suggests: the RTX 5090 figure is inference throughput through a separate accelerator, not the full served pipeline.
To try it, clone the GitHub repo, run hf download VAST-AI/EdenLive for the weights, and start with the 5B few-step config on the bundled samples; the live streaming service needs the 80 GB GPU.