This is the official model repository of the GenieRedux models, introduced by our works "Exploration-Driven Generative Interactive Environments" (CVPR'25) and "Learning Generative Interactive Environments By Trained Agent Exploration" (NeurIPS'24 D3S3 Workshop). GenieRedux is an open implementation of the Genie world model, while GenieRedux-G is its variant, adapted for virtual environments.
To download and use the models, please follow the instructions in our official code repository.
Our models are trained on the CoinRun environment, following the case study of the original work. We provide:
- Tokenizer - a spatial-temporal tokenizer with a VQ codebook;
- GenieRedux - Dynamics+LAM, with action embedding from the LAM;
- GenieRedux-G - Dynamics only, with one-hot action input.
In addition, we provide Agent57 checkpoints for generating test sets.
This model was created by INSAIT Institute, part of Sofia University, in Sofia, Bulgaria.