CycleGANs

Overview

From-scratch replication of CycleGAN for unpaired image-to-image translation. CycleGAN trains two generators (A→B and B→A) and two discriminators simultaneously, with a cycle consistency loss enforcing that translating an image and translating back recovers the original. This enables translation without paired training data. Based on Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks (Zhu et al., 2017).

Architecture

• Generator G: Domain A → Domain B (ResNet-based)
• Generator F: Domain B → Domain A (ResNet-based)
• Discriminators D_A, D_B: PatchGAN discriminators
• Losses: Adversarial + cycle consistency (L1) + identity loss

Training

• Dataset: Cityscapes (semantic segmentation maps ↔ street photos)
• Framework: PyTorch
• Generated images stored in output_images_val/

Paper

Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks — Zhu et al., 2017