🌓 Conditional GAN for Image Light Enhancement

A PyTorch implementation of a Conditional Generative Adversarial Network (cGAN) for adjusting image lighting conditions. This model can convert normal-light images to either low-light or bright-light conditions using controllable conditioning.

Original Image	Low Light Image	Bright Light Image

✨ Features

Feature	Description
🔄 Bidirectional Adjustment	Convert normal → bright or normal → low-light
🎚️ Conditional Control	Parameter range (-1.0 to 1.0) for light intensity
📊 Training Data	LOL-v2 dataset (Real and Synthetic)
🚀 Versatile	Supports both training and inference modes

Dataset Structure

The project uses the LOL-v2 dataset with the following structure:

dataset/LOL-v2/
    Real_captured/
        Train/
            Normal/
            Low/
            Bright/
        Test/
            Normal/
            Low/
            Bright/
    Synthetic/
        Train/
            Normal/
            Low/
            Bright/
        Test/
            Normal/
            Low/
            Bright/

To get the bright images run the generate_bright.py

python3 generate_bright.py

Then modify the image names by running remove_prefix.py

python3 remove_prefix.py

Usage

🛠️ Installation

1. PyTorch (with CUDA)

pip install torch torchvision --index-url https://download.pytorch.org/whl/cu124

2. All other dependencies

pip install -r requirements.txt

Training

To train the model from scratch:

python3 train_cgan.py

🔍 Testing

Convert single image with specific condition

# Convert to bright image
python3 test_cgan.py --input path/to/normal/image.jpg --output output/bright_image.jpg --condition 1.0

# Convert to low-light image
python3 test_cgan.py --input path/to/normal/image.jpg --output output/low_image.jpg --condition -1.0

✨ OR ✨

Convert to both conditions at once

python3 single_image_convert.py --input sample/lion.jpeg --output-dir output --model checkpoints/best_model.pth

Model Architecture

Generator

• Encoder-decoder architecture with skip connections
• Input: Normal-light image + condition value
• Output: Generated image with adjusted lighting

Discriminator

• Convolutional neural network
• Input: Image + condition value
• Output: Real/Fake prediction

Checkpoints

The model saves two types of checkpoints in the checkpoints directory:

• latest_checkpoint.pth: Latest training state
• best_model.pth: Best performing model based on generator loss