VAST Unveils EdenLive — Real-Time Generative Rendering

VAST (Project Eden) has unveiled EdenLive, a generative renderer that converts a rough proxy stream of a game scene, engine, or recording into finished video in real time, keeping world logic and state on the engine side.

Compiled by Sergey KostenchukPublished 2026-10-19Updated 2026-10-20

2026-10-19 Multimodal NVIDIA

🎨 VAST Unveils EdenLive — Real-Time Generative Rendering

A rough proxy stream from a 2D/3D scene, engine, or recording is converted into finished video on the fly: the engine holds the world state, while a neural network draws the image—a DINOv3 encoder feeds frame features into a diffusion transformer. Branches: Wan2.1-I2V-14B (480p) and Wan2.2-TI2V-5B (720p).

🌍 The separation of “engine holds state—neural network draws” frees up the entire real-time parameter budget for image quality, while health bars, cooldowns, and NPC states are not lost, as in action-conditioned world models. Open weights provide a ready-made base for game previews and video stylization.

👤 Weights (~210 GB) are available on Hugging Face, and code is on GitHub under Apache 2.0. The 5B model delivers 720p (~24.7 FPS) on a single RTX 5090. Real-time performance requires Linux and an NVIDIA GPU (80 GB, CUDA 13), and the FLUX component is non-commercial only.

Source 1: https://vast-eden.ai/?p=edenlive Source 2: https://huggingface.co/VAST-AI/EdenLive

Sources