Progressive Growing CycleGAN For High-Resolution Image Translation

Welcome to the Progressive Growing CycleGAN repository! This project explores an integration of CycleGAN with Progressive Growing GANs for high-resolution image-to-image translation. By combining these methods, we aim to achieve superior translation stability and fidelity when working with unpaired datasets.

Overview

This project aims to transform CIFAR-10 images into Claude Monet-inspired artworks by leveraging Progressive Growing CycleGAN, a method that incrementally scales image resolution and addes layers during training. The progressive approach enables the model to first capture coarse features and gradually refine high-resolution details, ensuring smoother, Monet-like textures and more stable training.

Features

• Progressive Resolution Scaling: The model starts training at a low resolution (e.g., 16x16) and gradually increases to 512x512, allowing stable learning at each stage.
• Improved Image Fidelity: Enhanced generation quality with smoother transitions, natural brushstroke textures, and coherent Monet-like aesthetics.
• Comprehensive Evaluation:
  1. Quantitative metrics:
  • Fréchet Inception Distance (FID)
  • Inception Score (IS)
  2. Qualitative visual comparisons.
• Baseline Comparison: Compared against the standard CycleGAN model for a fair evaluation.

Datasets

1. Source Domain: CIFAR-10 dataset, which contains diverse classes of natural images.
2. Target Domain: Claude Monet's artworks, sourced from UC Berkeley's official CycleGAN datasets.

Both datasets are resized dynamically during training to match the current resolution stage.

Key Results

• Quantitative: The proposed method achieves lower FID, indicating better alignment with the Monet style.

Model	FID	Inception Score
Progressive Growing CycleGAN	270.42	4.41
Standard CycleGAN	273.09	4.43

• Qualitative: Outputs are visually more refined and stylistically accurate.

Installation

1. Clone the repository:

   git clone https://github.com/your-repo-name/Progressive-Growing-CycleGAN.git
   cd Progressive-Growing-CycleGAN

2. Create a Python virtual environment and Install dependencies:

   make env

Repository Structure

Progressive_Growing_CycleGAN/
├── README.md
├── requirements.txt
├── notebook/
│   ├── Progressive_Growing_CycleGAN.ipynb
│   ├── Standard_CycleGAN.ipynb
|   ├── main.ipynb
├── lib/
│   ├── models.py
│   ├── losses.py
│   ├── datasets.py
│   ├── train.py
│   ├── utils.py
├── .github/workflows
|   ├── ci.yml
├── data/monet
|   ├── *.jpg
├── .gitignore
├── Makefile
├── Report.pdf
├── img
|   ├── *.jpg

Limitations and Future Work

• High Computational Demand: Progressive resolution scaling requires significant computational resources for higher resolution images.
• Source Domain Compatibility: Starting with datasets more visually aligned with Monet's thematic style could further enhance performance.

Future directions include:

• Optimizing computational efficiency with adaptive resolution scheduling.
• Experimenting with higher resolutions (beyond 512x512) for even finer outputs.

References

1. Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017.

2. Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. CycleGAN datasets. https://efrosgans.eecs.berkeley.edu/cyclegan/datasets/, 2017a.

3. Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. arXiv preprint arXiv:1703.10593, 2017b.

Citation

If you use this code or method in your work, please cite the report:

@misc{niu2024progressivecyclegan,
  author       = {Mu Niu and Hongyi Duan},
  title        = {Progressive Growing CycleGAN for High-Resolution Image Translation},
  year         = {2024}
}