Jan–Mar 2025

Conditional GAN Image-to-Image Translation

An image-to-image translation model that turns satellite imagery into map-like representations. I implemented a pix2pix-style conditional GAN: a U-Net generator that preserves spatial structure, and a PatchGAN discriminator that judges local realism rather than a single global score.

• Python
• PyTorch
• GANs
• Computer Vision

Highlights

• U-Net generator with skip connections to keep geographic structure aligned between input and output.
• PatchGAN discriminator for local texture discrimination instead of a single real/fake logit.
• GPU training loop in PyTorch with paired satellite/map examples.
• Quantitative evaluation using PSNR and SSIM alongside visual inspection of generated maps.