Generating-art-from-photographs

Implementation of CycleGAN in PyTorch and trained on Monet-Photo dataset

CycleGAN

CycleGAN imposes a constraint that the model should be cycle consistent. If G : X -> Y transforms an image from domain X to target (Y) distribution and F : Y -> X transforms an image from domain Y to target (X) distribution, then cycle consistency constraint ensures that the reconstruction error is minimized i.e F(G(x)) ~ x.

We also want the color distribution to be preserved, so we also impose an identity loss. It says that the functions G and F must be able to represent the Identity function. So G(y) ~ y which ensures that the colors are not drastically changed when not necessary.

Key elements of the model

• Instance Normalization
• Residual Blocks

Some Training images

Photo images

Monet images

Visualizing Training stages

                                               Original Input Monet Images

                                               Original Input Photographs

                                                Epoch 0 Predicted photos

                                                 Epoch 50 Predicted photos

                                                   Epoch 100 Predicted photos

                                               Epoch 0 Reconstructed monets

                                                 Epoch 50 Reconstructed monets

                                                 Epoch 100 Reconstructed monets

                                                 Epoch 0 Predicted monets

                                                  Epoch 50 Predicted monets

                                                   Epoch 100 Predicted monets

                                                Epoch 0 Reconstructed photos

                                                 Epoch 50 Reconstructed photos

                                                  Epoch 100 Reconstructed photos

Plotting losses over time

Final generated Portraits from Photographs