FaceForge

Sketch-to-Face Generation with StyleGAN2

This project implements a deep learning pipeline that converts sketch images into realistic face images using a pretrained StyleGAN2 generator and a custom OME Encoder. The encoder learns to predict latent style modifications from a sketch input to drive face synthesis.

---

Overview

• Model Architecture:

  • OME Encoder: A ResNet-based encoder that takes a sketch and previous face as input and outputs latent delta_w vectors.
  • StyleGAN2: A pretrained generator (frozen) used for high-fidelity face synthesis.

• Losses Used:

  • L2 Loss: Pixel-level similarity.
  • LPIPS Loss: Perceptual similarity.
  • Identity Loss: Preserves identity using FaceNet features.

• Dataset: CelebA-HQ 256x256 Resized

---