Tripo releases EdenLive to render game prototypes from reference video

Tripo’s VAST AI Research announced EdenLive on October 18, releasing a generative renderer for game prototypes. It takes reference footage, a target first frame and a text prompt, then generates a detailed or stylized video stream. Developers can explore a prototype’s appearance while its code continues to run the game rules.

The model release includes 5B and 14B weights, each with offline bidirectional and streaming variants. The code repository also provides a workbench and EdenGolem v0.1 for building a playable greybox game.

Game code keeps collisions, scores and outcomes

The reference stream carries layout, camera movement and action timing. Appearance comes from the target first frame and prompt. In the official fighting example, game code calculates damage and updates health while EdenLive generates the visible frames.

EdenGolem’s documented scope covers standalone Three.js prototypes: a silent desktop game, keyboard and mouse controls, one continuous round, and reachable win and loss states. Its workflow covers the brief, scene, behavior and final acceptance, including checks with real inputs. It does not promise a complete commercial game.

This could help developers compare visual directions for an existing playable loop. Persistent rules remain inspectable in the upstream program, while generated details can still drift. VAST says appearances may become inconsistent when a character or location returns after a long absence from view.

The RTX 5090 figure comes from an optimized profile

The LightX2V guide reports about 24.7 inference FPS at 1280 × 704 for the 5B autoregressive model on one RTX 5090. That profile uses INT8 quantization and streaming decoding. Its timing excludes model loading, compilation, warmup and final video saving; it is not a measurement of complete player-input-to-display latency.

The main README lists Linux, CUDA 13 and an 80 GB NVIDIA GPU as its verified environment. The RTX 5090 result belongs to the separate optimized path. The complete weight download is about 210 GB, and workbench game generation requires an authenticated Claude CLI.

Model components have different licences

First-party code is Apache 2.0. The model card lists Apache 2.0 for several weight components, MIT for CLIP and a separate DINOv3 licence for the vision encoder.

The workbench’s first-frame generator uses FLUX.2 [klein] 9B under the FLUX Non-Commercial License. The project documents that workbench as non-commercial and non-production. Its offline and real-time video pipelines do not load that FLUX component. Deployment decisions therefore need to follow the components actually used, rather than treating the entire release as Apache 2.0.