Pix2Pix

Overview

From-scratch replication of Pix2Pix for paired image-to-image translation. Unlike CycleGAN, Pix2Pix requires paired training data and uses a U-Net generator with a PatchGAN discriminator that classifies overlapping 70×70 image patches as real or fake. Based on Image-to-Image Translation with Conditional Adversarial Networks (Isola et al., 2017).

Architecture

Generator: U-Net with skip connections (encoder-decoder + lateral connections at each resolution)

Discriminator: PatchGAN — classifies 70×70 patches rather than the full image, encouraging sharp local texture

• Combined adversarial + L1 reconstruction loss
• L1 weight encourages low-frequency correctness

Training

• Dataset: Aerial2Map (aerial photography ↔ map tiles)
• Framework: PyTorch

Paper

Image-to-Image Translation with Conditional Adversarial Networks — Isola et al., 2017