Pix2PixHD: Edges to Faces - CelebA-HQ

(

Generating photorealistic faces from edge maps using Conditional GANs (Pix2PixHD)

🌟 Overview

Welcome to the Pix2PixHD: Edges to Faces project! In this repository, I’ve implemented and trained a modified Pix2PixHD model to generate high-resolution (1024x1024) photorealistic face images from edge maps using the CelebA-HQ dataset. This project extends the groundbreaking work presented in “High-Resolution Image Synthesis and Semantic Manipulation with Conditional GANs”, providing an exciting tool for tasks such as sketch-based image synthesis and image editing.

🚀 Results

The model takes edge maps as input and generates stunning, realistic face images. Below are some results from Phase 1 of training:

Input (Edge Map)	Output (Generated Image)

These outputs are produced using Global Generator with 512x512 resolution, giving exceptional detail and realism!

📄 Original Paper

This implementation is based on the paper “High-Resolution Image Synthesis and Semantic Manipulation with Conditional GANs” by Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, and others. Pix2PixHD builds on the standard Pix2Pix architecture to produce images with higher resolutions and fidelity, especially useful for tasks such as semantic image synthesis.

📂 Project Structure

• /data: Contains the dataset preprocessing scripts and edge map generation using Canny edge detection.
• ./: The core PyTorch implementation of Pix2PixHD, including Global Generator, Local Enhancer, and the multi-scale Discriminators.
• /output: The generated outputs from the model.
• train.py: Script for training the model on edge-to-image task.

⚙️ How to Run

Prerequisites

To get started, make sure you have the following installed:

• Python 3.12.x
• PyTorch 2.4.x with CUDA

Installation

1. Clone the repository:

   git clone https://github.com/inventwithdean/pix2pixhd.git
   cd pix2pixhd

2. Download the CelebA-HQ dataset and place it in the /data folder. You can find it here.
3. Make sure you have all the faces inside data/faces folder. And create a directory data/sketches.
4. Generating edge maps:

   python create_dataset.py

Training

To train the model on the edge-to-face task, run:

python train.py

Create directories enhanced_checkpoints and global_checkpoints for checkpointing.

📊 Features

• High-Resolution Synthesis: Generates 1024x1024 images from edge maps.
• Conditional GAN Architecture: Uses multi-scale discriminators for improved realism.
• Instance-Level Control (WIP): Planned updates for fine-tuning specific features in generated images.

📝 Future Work

• Instance-wise control for fine-grained editing of facial features.
• Further improvements to training strategies for faster convergence.
• Integration of new datasets for generalization beyond CelebA-HQ.

🛠️ Contributing

Feel free to fork this repo, submit issues, or make pull requests. Contributions are always welcome to improve the codebase!

📬 Contact

If you have any questions or feedback, feel free to reach out to me via LinkedIn or open an issue on GitHub.

💡 Acknowledgements

This work is inspired by the original Pix2PixHD implementation by NVIDIA and the creators of the CelebA-HQ dataset.

---

Enjoy experimenting! If you like the project, don’t forget to give it a ⭐ on GitHub!