Follow instructions in Boris Dayma's DALL-E mini repo to download the latest DALL-E Mega model.

This Dockerfile will run the notebooks from that repo and generate you some images. Just like the Huggingface space does. Haven't tried training yet.

On an RTX 3090, it takes about 5.6 seconds per image (so 50 seconds for a 9-pack of images).

Image generation also worked on CPU, but took several minutes per image.

The images look great:

docker build -t dallemega .

Replace YOUR_PATH_HERE with wherever wandb downloaded your dall-e checkpoint to:

docker run --gpus all -it --rm -v $(realpath ~/notebooks):/notebooks -v /YOUR_PATH_HERE/artifacts/:/artifacts -p 8888:8888 dallemega

That will launch a Jupyter notebook instance, open a browser to localhost:8888/?token=... and copy the token from stdout.

If your docker run command fails, you may need to install nvidia-container-toolkit. That's an OS-specific process so... Google around. Good luck.

Make a notebook in /notebooks and copy/paste stuff into it from the dalle-mini inference notebook.

Replace the DALLE_MODEL path with yours. So if you have checkpoint 15, it looks like:

DALLE_MODEL = "/artifacts/mega-1:v15"