kaiyuyue/sphere-encoder-fid-artifacts

Sphere Encoder FID Evaluation Artifacts

This repository contains the evaluation artifacts for the paper Image Generation with a Sphere Encoder.

Project Page | GitHub Repository

These artifacts include data statistic files (fid_stats) and reference images (fid_refs) used to calculate Fréchet Inception Distance (FID) for generative models across several datasets, including CIFAR-10, ImageNet, Animal Faces, and Oxford Flowers.

Workspace Setup

Download the evaluation artifacts and place them in your ./workspace/ directory. The directory tree should look like this:

./workspace/
├── fid_stats
    |── fid_stats_extr_animal-faces_256px.npz
    |── fid_stats_extr_cifar-10_32px.npz
    |── fid_stats_extr_flowers-102_256px.npz
    |── fid_stats_rand-50k_imagenet_256px.npz
├── fid_refs
    |── ref_images_imagenet_256px/images

Sample Usage

To evaluate a trained model using these artifacts, you can use the evaluation script provided in the GitHub repository:

./run.sh eval.py \
  --job_dir sphere-base-base-cifar-10-32px \
  --forward_steps 1 4 \
  --report_fid rfid gfid \
  --use_cfg True \
  --cfg_min 1.2 \
  --cfg_max 1.2 \
  --cfg_position combo \
  --rm_folder_after_eval True

Citation

@article{yue2025image,
  title={Image Generation with a Sphere Encoder},
  author={Yue, Kaiyu and Jia, Menglin and Hou, Ji and Goldstein, Tom},
  journal={arXiv preprint arXiv:2602.15030},
  year={2025}
}