Generative AI

Sketch-to-Photo Generation using GANs

Converting face sketches into realistic photos with a pix2pix-style GAN.

2025

OVERVIEW

The Project

A deep-learning pipeline that transforms human face sketches into full-colour, photo-realistic images using a Generative Adversarial Network. The generator maps sketches to RGB photos while the discriminator refines realism, trained on paired sketch-photo datasets in PyTorch.

Objectives

Train a GAN to reconstruct photo images from sketch input.

Use L1 loss + adversarial feedback for better image fidelity.

Generalize across validation images with normalized training data.

Tools & Technologies

PyTorchTorchvisionMatplotlibPILCUDA

METHODOLOGY

The Approach

Pair grayscale sketches with their RGB photos via a custom dataset loader.

Generator: encoder–decoder CNN (3 conv + 3 deconv, BatchNorm, ReLU, Tanh).

Discriminator: CNN with LeakyReLU on the 4-channel sketch+photo input.

Train with BCE adversarial loss + L1 pixel loss (Adam), 100 epochs.

OUTCOME

Results & Learnings

Generated realistic faces capturing structure, contours and lighting.

Side-by-side per-epoch visualization tracked convergence.

Key Learnings

Combining pixel-based and adversarial objectives sharpens output.
Conditioned GAN training and data normalization are key to fidelity.

Related Case Studies