Few-shot image generation is a challenging task since it aims to generate diverse new images for an unseen category with only a few images. Existing methods suffer from the trade-off between the quality and diversity of generated images. To tackle this problem, we propose Hyperbolic Attribute Editing (HAE), a simple yet effective method. Unlike other methods that work in Euclidean space, HAE captures the hierarchy among images using data from seen categories in hyperbolic space. Given a well-trained HAE, images of unseen categories can be generated by moving the latent code of a given image toward any meaningful directions in the Poincaré disk with a fixing radius. Most importantly, the hyperbolic space allows us to control the semantic diversity of the generated images by setting different radii in the disk. Extensive experiments and visualizations demonstrate that HAE is capable of not only generating images with promising quality and diversity using limited data but achieving a highly controllable and interpretable editing process.

Official implementation of HAE for few-shot image generation. Our code is modified from pSp.

• Linux
• NVIDIA GPU + CUDA CuDNN (CPU may be possible with some modifications, but is not inherently supported)
• Python 3

• Clone this repo:

git clone https://github.com/lingxiao-li/HAE.git
cd HAE

• Dependencies:
  We recommend running this repository using Anaconda. All dependencies for defining the environment are provided in /HAE/hae/environment/hae_env.yaml.

Then download the pre-trained models we provide from Google Drive

Then put the pre-trained models under /HAE/hae/pretrained_models

Related datasets can be downloaded at: https://github.com/bcmi/Awesome-Few-Shot-Image-Generation

After downloading the pre-trained models and datasets, change the corresponding path at /Codes/hae/configs/paths_config.py

Go to the path to HAE:

cd /PATH_TO/HAE/hae

Then you can train the model using:

python scripts/train.py \
--dataset_type=flowers_encode_eva \
--psp_checkpoint_path=/PATH_TO/psp_flowers.pt \
--exp_dir=OUTPUT_PATH \
--feature_size=512 \
--workers=8 \
--batch_size=8 \
--test_batch_size=8 \
--test_workers=8 \
--val_interval=80000 \
--save_interval=5000 \
--encoder_type=GradualStyleEncoder \
--start_from_latent_avg \
--lpips_lambda=1 \
--l2_lambda=1 \
--image_interval=1000 \
--hyperbolic_lambda=0.3 \
--reverse_lambda=1

Inference the images using:

python scripts/inference.py \
--exp_dir=OUTPUT_PATH \
--checkpoint_path=/PATH_TO/hae_flowers.pt \
--data_path=PATH_TO/flowers_eva/test \
--test_batch_size=4 \
--test_workers=4

Please see the comments in the notebook:

Interpolation & Perturbation.ipynb and UMAP_Visualization.ipynb

If you use this code for your research, please cite our paper The Euclidean Space is Evil: Hyperbolic Attribute Editing for Few-shot Image Generation:

@InProceedings{Li_2023_HAE,
    author    = {Li, Lingxiao and Zhang, Yi and Wang, Shuhui},
    title     = {The Euclidean Space is Evil: Hyperbolic Attribute Editing for Few-shot Image Generation},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2023},
    pages     = {22714-22724}
}