Conditional GAN:

This project implements a Cycle-Consistent Generative Adversarial Network (CycleGAN) to convert Monet-style paintings into realistic photographs using the MonetGAN dataset. Unlike traditional GANs, CycleGAN does not require paired image data. Instead, it learns to perform unpaired image-to-image translation by using cycle consistency and adversarial training.

• Platform: Google Colab
  
  

• Dataset: MonetGAN (Unpaired Monet paintings and real photos)
  
  

 Architecture:

 Generator Network (G_AB & G_BA)

• Converts images between the two domains: Monet → Photo and Photo → Mone

• Uses:

  • Reflection padding

  • Downsampling with convolutions

  • Residual blocks for stable translation

  • Upsampling to return to original resolution

  • Tanh activation to produce image outputs
    
    

 Discriminator Network:

• PatchGAN structure (classifies image patches as real or fake)
  
  

• Works separately for each domain
  
  

• Outputs a probability map instead of a single value
  
  

 Loss Functions

• Adversarial Loss: Encourages generators to produce realistic images to fool the discriminators
  
  

• Cycle Consistency Loss: Ensures image translations are consistent when cycled back to original domain
  
  

• Identity Loss: Ensures the generator preserves color/content when input image is already in target domain
  
  

 Training Process:

• Trained for 100 epochs

• Used unpaired datasets of Monet paintings and real photos

• Images were preprocessed to 256×256 resolution

• Optimized using the Adam optimizer
  
  

Results

The model successfully learned to generate realistic photos from Monet-style paintings.

📸 Sample Output

• Input: Monet Painting                                                                               Output: Realistic Photo Version

Google colab Notebook page :