CycleGAN: Unpaired Image-to-Image Translation 🦓🐴

implementation of CycleGAN for unpaired image-to-image translation. The included dataset is the classic horse2zebra dataset.

Architecture

CycleGAN consists of:

• 2 Generators: G_AB (A to B) and G_BA (B to A)
• 2 Discriminators: D_A (judges domain A) and D_B (judges domain B)

Quick Start

1. Set up Python environment:

   python -m venv venv
   source venv/bin/activate  # On Windows: venv\Scripts\activate

2. Install dependencies:

   pip install -r requirements.txt

3. Train the model (requires GPU for reasonable training time):

   python scripts/train.py

   • Checkpoints saved every epoch in checkpoints/
   • Training for 200 epochs (takes several hours on GPU)

4. Generate translations:

   python scripts/test.py

   • Translated images saved in outputs/
   • Uses the final model checkpoint by default

Example Results

Here are some side-by-side comparisons of the translations from horses (domain A) to zebras (domain B) using the epoch 10 checkpoint (cherry-picked):

Example 1: n02381460_800.jpg

Original (Horse)	Translated (Zebra)

Example 2: n02381460_1300.jpg

Original (Horse)	Translated (Zebra)

Example 3: n02381460_470.jpg

Original (Horse)	Translated (Zebra)

Notes

CycleGAN uses three types of losses:

• adversarial loss
  • makes generated images look realistic
  • discriminator can't tell real from fake
• cycle consistency loss
  • A → B → A ≈ A (most important)
  • preserves content while changing style

identity loss

• G_AB(B) ≈ B (optional but helpful)
• prevents unnecessary color changes

training process

1. generator creates fake images
2. discriminator learns to identify fakes
3. generator improves to fool discriminator
4. cycle consistency ensures meaningful translations
5. iterate until convergence

patchGAN discriminator

• classifies 70×70 patches instead of whole images
• more efficient and effective for textures

instance normalization

• better than batch normalization for style transfer
• normalizes each image independently

replay buffer

• stores past generated images
• prevents discriminator from overfitting

generator architecture

• ResNet-based with skip connections
• encoder-decoder structure
• 9 residual blocks by default

discriminator architecture

• PatchGAN: classifies overlapping patches
• produces a grid of real/fake predictions
• more parameter-efficient than full image

training details

• Adam optimizer (lr=0.0002, β1=0.5)
• linear learning rate decay after 100 epochs
• batch size of 1 for stability

References

• Original CycleGAN Paper
• Project Page
• Official Implementation