Oasis is an interactive world model developed by Decart and Etched. Based on diffusion transformers, Oasis takes in user keyboard input and generates gameplay in an autoregressive manner. We release the weights for Oasis 500M, a downscaled version of the model, along with inference code for action-conditional frame generation.
For more details, see our joint blog post to learn more.
And to use the most powerful version of the model, be sure to check out the live demo as well!
git clone https://github.com/etched-ai/open-oasis.git
cd open-oasis
# Install pytorch
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121
# Install other dependencies
pip install einops diffusers timm av
huggingface-cli login
huggingface-cli download Etched/oasis-500m oasis500m.pt # DiT checkpoint
huggingface-cli download Etched/oasis-500m vit-l-20.pt # ViT VAE checkpoint
We include a basic inference script that loads a prompt frame from a video and generates additional frames conditioned on actions.
python generate.py
The resulting video will be saved to video.mp4. Here's are some examples of a generation from this 500M model!
Hint: try swapping out the
.mp4input file in the script to try different environments!