vast-ai-research/edenlive is a real-time generative renderer that builds frames for interactive worlds from reference video and sentence instructions. State changes in the world are handled by a coding agent or an engine, while EdenLive renders the look on screen. No version information is given in FACTS for this release, and the repository description presents an architecture that separates state handling from rendering.
In this architecture, the engine connects the reference video it produces to a target first frame and to sentence instructions to build a detailed screen. It is licensed under Apache-2.0 and publishes 5B and 14B weights on Hugging Face. The public items include weights for bidirectional inference and few-step inference, real-time streaming inference code, a game skill for manipulating sentence instructions, and a frontend to check the whole flow. The project page presents 480P and 720P output and playback beyond 24 frames per second, and also covers a 720P single-GPU batch setup for the 5B model.
The documented verification environment, however, is Linux x86-64 with an 80GB NVIDIA GPU and Python 3.10. The figure of about 24.7 fps at 1280x704 on a single RTX 5090 is limited to the table in the LightX2V integration section. So it is more accurate to read it as a result presented under the published environment and integration conditions rather than as typical consumer graphics performance.