Nested Diffusion Processes for Anytime Image Generation

Noam Elata, Bahjat Kawar, Tomer Michaeli, and Michael Elad, Technion - Israel Institute of Technology.

🔗 Project Webpage | 🤗 Huggingface Demo | ArXiv

Preparation

Please refer to environment.yml for packages that can be used to run the code.

Pretrained models will download automatically upon running the relevant scripts.

Sampling from Nested Stable Diffusion

This code implements Nested Diffusion, used with Stable Diffusion v1-5 and the Diffusers library.

Image generation can be run with the following script:

python sample_text_to_image.py --outer <number of outer steps> --inner <number of inner steps>  \
               --outdir samples --prompt "<your text prompt>"

Default parameters reproduce the image shown at the top.

To run Nested Diffusion with accelerated DPM-Solver++ inner diffusion scheduler, please add the --dpm-solver argument to the script. It is possible to use fewer inner steps with DPM-Solver++.

If you have less than 10GB of GPU RAM available please add the --fp16 argument to the script.

ImageNet Generation with Nested Diffusion

This code implements Nested Diffusion, used with DiT as a pretrained model.

Image generation with Nested Diffusion can be run with the following script:

python generate_imagenet.py --steps <number of outer steps> --nested <number of inner steps> \
               -n <number of samples to generate> -b <batch size> \
               -o <output directory path> --cfg <value of CLF to use>

For Vanilla DiT diffusion process please run with the following arguments:

python generate_imagenet.py --steps <number of vanilla steps> \
               -n <number of samples to generate> -b <batch size> \
               -o <output directory path> --cfg <value of CLF to use>

This script will save all intermediate predictions in separate directories.

For multi-GPU sampling or ddim sampling, please see the usage of generate_imagenet.py.

References and Acknowledgements

@InProceedings{Elata_2024_WACV,
    author    = {Elata, Noam and Kawar, Bahjat and Michaeli, Tomer and Elad, Michael},
    title     = {Nested Diffusion Processes for Anytime Image Generation},
    booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
    month     = {January},
    year      = {2024},
    pages     = {5018-5027}
}