VAST, the company behind Tripo, open-sourced EdenLive on GitHub and Hugging Face. A game engine or coding agent runs the rules and world state and outputs a rough reference video; EdenLive renders it in real time into detailed frames, guided by a target first frame and a text prompt. VAST released 5B and 14B weights in bidirectional and few-step autoregressive variants, real-time streaming inference code, and a skill that turns a text prompt into a playable game prototype. Per VAST, the project page shows over 24 FPS, minute-long generation and 480P/720P output; with LightX2V the 5B model reaches about 24.7 FPS at 1280x704 on one RTX 5090.
Key Takeaways
- ✓The idea is to split state from appearance: an upstream program or engine keeps the rules and world state, and EdenLive only renders the reference video into detailed frames.
- ✓Released: 5B and 14B weights (bidirectional and few-step chunk-AR), real-time streaming inference code, the game-prototype-turbo skill and a frontend workbench.
- ✓Per VAST: over 24 FPS, minute-long generation, 480P/720P output; with LightX2V the 5B model reaches about 24.7 FPS at 1280x704 on one RTX 5090.
- ✓Built on Wan2.1-I2V-14B (480p) and Wan2.2-TI2V-5B (720p); the full weight tree is about 210 GB, and the verified setup is an 80 GB NVIDIA GPU with CUDA 13.
- ✓Licensing caveat: first-party code is Apache 2.0, but the workbench first-frame model FLUX.2 [klein] 9B is non-commercial, and the DINOv3 encoder has its own license.
Key Decision Metrics at a Glance
Heavy Claude Code use: compare subscription limits and API bills
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Core Background & Industry Pain Points
Interactive world models usually ask one video model to handle both how the world changes and how it looks. Rules drift, long runs get unstable, and most systems need datacenter GPUs. For indie game developers, the useful case is real-time output on a consumer card with game logic that stays under control.
Architecture Highlights & Mechanism
EdenLive splits the two jobs. An upstream game engine or coding agent runs the rules and keeps the world state, outputting a rough reference video such as greybox footage. EdenLive uses that video as the interface and, guided by a target first frame and a text prompt, renders detailed frames in real time. VAST ships a 5B model (Wan2.2-TI2V-5B, 720p) and a 14B model (Wan2.1-I2V-14B, 480p), each in bidirectional and few-step chunked autoregressive variants, with streaming over WebRTC.
Benchmarks & Measured Results
VAST has not published standardized comparisons against other world models. All checkable numbers are vendor figures: the project page shows over 24 FPS, minute-long generation and 480P/720P output; with LightX2V quantization, operator fusion and CUDA Graphs, the 5B model reaches about 24.7 FPS at 1280x704 on a single RTX 5090, enough for 24 FPS playback.
Developer Quick Start
Code is at GitHub VAST-AI-Research/EdenLive and weights download with hf download VAST-AI/EdenLive (about 210 GB in total). The verified setup is Linux, an 80 GB NVIDIA GPU, CUDA 13 and Python 3.10. Two greybox samples ship for offline inference. The prompt-to-game workbench needs the claude CLI installed and logged in, and its first-frame model FLUX.2 [klein] 9B is non-commercial, so check each component license before commercial use.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.