TotoRoam
Envision
CompSoc
Abstract
This project implements CycleGAN : a framework for image-to-image translation that requires no paired training data. The core idea is to learn two mappings simultaneously: G : X → Y and F : Y → X, and enforce a cycle consistency constraint such that F(G(x)) ≈ x and G(F(y)) ≈ y. This removes the need for aligned image pairs, which are expensive or impossible to collect for many tasks.
The task chosen for this implementation is artistic style transfer, specifically translating real photographs into a target artistic style (Ghibli-style anime). The generator uses a sequential encoder-decoder with a residual block bottleneck, reflection padding and instance normalisation, the discriminator uses a 70×70 PatchGAN, and the training loop implements identity loss, adversarial loss, and cycle consistency loss as per the original paper.
Abstract
This project implements CycleGAN : a framework for image-to-image translation that requires no paired training data. The core idea is to learn two mappings simultaneously: G : X → Y and F : Y → X, and enforce a cycle consistency constraint such that F(G(x)) ≈ x and G(F(y)) ≈ y. This removes the need for aligned image pairs, which are expensive or impossible to collect for many tasks.
The task chosen for this implementation is artistic style transfer, specifically translating real photographs into a target artistic style (Ghibli-style anime). The generator uses a sequential encoder-decoder with a residual block bottleneck, reflection padding and instance normalisation, the discriminator uses a 70×70 PatchGAN, and the training loop implements identity loss, adversarial loss, and cycle consistency loss as per the original paper.
Report Information
Report Details
Created: May 17, 2026, 11:14 p.m.
Approved by: None
Approval date: None
Report Details
Created: May 17, 2026, 11:14 p.m.
Approved by: None
Approval date: None