Awesome
Adaptive-Feature-Interpolation-for-Low-Shot-Image-Generation (ECCV 2022)
Paper: https://arxiv.org/abs/2106.10777
Below are generated images trained on datasets with 60-1000 samples. Fast access to datasets: https://github.com/odegeasslbc/FastGAN-pytorch (Liu's repo).
<p align="center"> <img src="comp_gen.jpg" align="middle" width="600"> </p> <p align="center"> (Left: StyleGAN2-ADA; Right: + Adaptive Feature Interpolation) </p>Usage:
Adaptive Feature Interpolation
Create a batch of new features from a batch of old features:
new_feature = near_interp(old_feature, k, augment_prob)
where k, augment_prob can be generated by function dynamic_prob or defined by user. Please refer to interp_feature.py for more details. Example of implementation in StyleGAN2 can be found in the corresponding folder.
Citation
@InProceedings{10.1007/978-3-031-19784-0_15,
author="Dai, Mengyu
and Hang, Haibin
and Guo, Xiaoyang",
title="Adaptive Feature Interpolation for Low-Shot Image Generation",
booktitle="Computer Vision -- ECCV 2022",
year="2022",
publisher="Springer Nature Switzerland",
address="Cham",
pages="254--270"
}