Official implementation of the paper ["DEff-GAN: Diverse Attribute Transfer for Few-Shot Image Synthesis " by Rajiv Kumar and G. Sivakumar] Arxiv: https://arxiv.org/pdf/2302.14533v1.pdf) Scitepress: https://www.scitepress.org/Papers/2023/117996/117996.pdf).
- python 3.5 or above
- pytorch 1.1.0 or above
pip install -r requirements.txt
The code was tested on an environment that can be imported using the environment.yml file.
conda env create -f environment.yml
However, there are lots of extra packages that may consume more disk space.
[NEW!] Our Colab tutorial is released!
To train a model with the default parameters from our paper run:
python train.py
Training one model on image with 128x128 image size and 6 stages should take about 1-2 hours on an NVIDIA GeForce GTX 1080Ti.
To affect sample diversity and image quality we recomment playing around with the learning rate scaling (default is 0.1) and the number of trained stages (default is 6 \ recommeded is '8').
This can be especially helpful if the images are more complex (use a higher learning rate scaling) or you want to train on images with higher resolution (use more stages).
For example, increasing the learning rate scaling will mean that lower stages are trained with a higher learning rate and can, therefore, learn a more faithful model of the original image.
Training on more stages can help with images that exhibit a large global structure that should stay the same.
The output is saved to TrainedModels/.
To sample more images from a trained model run:
This will use the model to generate num_samples images in the default as well as scaled resolutions.
The results will be saved in a folder Evaluation in the model_dir.
The default unconditional image generation is geared to also induce diversity at the edges of generated images. When generating images of arbitrary sizes (especially larger) this often break the image layout.
The folder Image contains the raw images that were used in paper.
This code implementation borrows heavily from implementation of the ConSinGAN paper. We thank the authors for sharing their code publicly!
If you found this code useful please consider citing:
@conference{visapp23,
author={Rajiv Kumar and G. Sivakumar},
title={DEff-GAN: Diverse Attribute Transfer for Few-Shot Image Synthesis},
booktitle={Proceedings of the 18th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications - Volume 5: VISAPP,},
year={2023},
pages={870-877},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0011799600003417},
isbn={978-989-758-634-7},
}