brownvc/R3GAN-ImgNet-32x32

R3GAN: ImageNet 32×32

The class-conditional (1,000 classes) R3GAN generator for ImageNet at 32×32, from the NeurIPS 2024 paper The GAN is dead; long live the GAN! A Modern GAN Baseline (the proposed configuration, config E).

Dataset	ImageNet-1k (ILSVRC 2012)
Resolution	32×32
Conditioning	class-conditional (1,000 classes)
FID (50k samples)	1.27, as reported in the paper
File	network-snapshot-final.pkl: the final training snapshot; its G_ema network is the generator
Code	https://github.com/brownvc/R3GAN
Licence	CC BY-NC 4.0

Generating images

git clone https://github.com/brownvc/R3GAN && cd R3GAN
python gen_images.py --seeds=0-7 --outdir=out --class=0 \
    --network=https://huggingface.co/brownvc/R3GAN-ImgNet-32x32/resolve/main/network-snapshot-final.pkl

The model is class-conditional, so --class is required (class 0 is tench). The code requires the same packages as the official StyleGAN3 repository.

Intended use and limitations

The model is intended for research on generative models. It reproduces the statistics of its training data, ImageNet-1k (1.28 million web images in 1,000 classes), including the known biases of a web-scraped dataset. Do not use it to create deceptive content or content that could cause harm.

Licence

The weights are released under CC BY-NC 4.0. The code that runs them is a separate work: the R3GAN/ directory of the GitHub repository is under the MIT licence, and the rest of the repository, which derives from StyleGAN3, is under NVIDIA's Source Code License.

Citation

@inproceedings{huang2024the,
  title={The {GAN} is dead; long live the {GAN}! A Modern {GAN} Baseline},
  author={Nick Huang and Aaron Gokaslan and Volodymyr Kuleshov and James Tompkin},
  booktitle={The Thirty-eighth Annual Conference on Neural Information Processing Systems},
  year={2024},
  url={https://openreview.net/forum?id=OrtN9hPP7V}
}

Acknowledgements

Brown University and Cornell University. Nick Huang was supported by a Brown University Division of Research Seed Award, James Tompkin by NSF CAREER 2144956, and Volodymyr Kuleshov by NSF CAREER 2145577 and NIH MIRA 1R35GM15124301. Compute was provided by Databricks Mosaic Research.