Official implementation for Long Tail Image Generation Through Feature Space Augmentation and Iterated Learning, accepted in LXAI CVPR2024 as an extended abstract.
The tested dataset is available here, and an example usage script is available here.
0_img_to_vector.py: Convert images to latent vectors.1_train_iterated.py: Run iterated training for the Encoder, Decoder and Classifier.2_fuse.py: Create class activation maps and fuse long tail images based on their nearest neighbors.3_vec_to_img.py: Convert fused vectors into new images.
The current code is very memory hungry, proceed with caution and start with few images.