cyclegan-keys-to-raisins

CycleGAN: Keys to Raisins

A CycleGAN model that converts images of keys into images of raisins (and vice versa) using unpaired image-to-image translation.

Overview

This project implements the classic CycleGAN architecture from Zhu et al., "Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks" (ICCV 2017).

The model learns to translate between two unpaired image domains:

• Domain A: Keys (metal objects with various head shapes, shafts, and teeth)
• Domain B: Raisins (small dark dried grapes on various surfaces)

Architecture

• Generator: ResNet-9 with 9 residual blocks, InstanceNorm, reflection padding
• Discriminator: 70x70 PatchGAN (C64-C128-C256-C512 → 1)
• Loss Functions:
  • Least-squares GAN loss for adversarial training
  • L1 cycle consistency loss (λ=10)
  • L1 identity loss (λ=0.5×cycle)

Training Details

Dataset

The training data was generated synthetically:

• 300 key images with varied head shapes (circle, double-circle, oval, square), shaft lengths, and tooth patterns
• 300 raisin images with organic blob shapes, varied dark colors, wrinkles, and highlights on different backgrounds
• 90/10 train/test split

Usage

Training

# Generate data and train
python train_cyclegan.py --epochs 200 --dataroot ./datasets/keys2raisins --output_dir ./outputs

Inference

# Key -> Raisin
python inference.py --input key_photo.jpg --checkpoint G_AB_final.pth --output raisin.jpg

# Raisin -> Key
python inference.py --input raisin_photo.jpg --checkpoint G_BA_final.pth --output key.jpg

Python API

import torch
from inference import Generator, load_generator

device = 'cuda' if torch.cuda.is_available() else 'cpu'

# Load key -> raisin generator
G = load_generator('G_AB_final.pth', ngf=64, n_residual_blocks=9, device=device)

# Your image should be 256x256 RGB, normalized to [-1, 1]
# The generator returns a translated tensor

Files

Citation

@inproceedings{zhu2017unpaired,
  title={Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks},
  author={Zhu, Jun-Yan and Park, Taesung and Isola, Phillip and Efros, Alexei A},
  booktitle={ICCV},
  year={2017}
}