How aiff22/dped works
This repository implements a TensorFlow photo-enhancement workflow: train camera-specific models on paired mobile/DSLR images, then apply trained or provided checkpoints to test photos. The training and data/model internals are supported by sampled source; inference capabilities and outputs are described in the README, but test_model.py was not sampled, so its internal wiring is represented only where documented.
Main entry points: train_model.py Train models · test_model.py Enhance test photos · vgg.py VGG-19 features · load_dataset.py Paired data loader
Components 15
Workflows
- Train models
train_model.py - Enhance test photos
test_model.py
Training Components
- Enhancement ResNet
models.py - Texture discriminator
models.py - VGG-19 features
vgg.py - Quality metrics
ssim.py - Training and image utilities
utils.py
Data and Models
- Paired data loader
load_dataset.py - DPED image pairs
dped - Pretrained VGG weights
vgg_pretrained - Saved checkpoints
models - Provided checkpoints
models_orig - Training previews
results - Enhanced test images
visual_results
Other components
- Engineer / user
Connections 17
- Engineer / user → Train models: starts training
- Engineer / user → Enhance test photos: requests enhancement
- Train models → Training and image utilities: parses options
- Train models → Paired data loader: loads batches
- Paired data loader → DPED image pairs: reads image pairs
- Train models → Enhancement ResNet: builds enhancement graph
- Train models → Texture discriminator: builds adversarial graph
- Train models → VGG-19 features: builds content features
- VGG-19 features → Pretrained VGG weights: loads weights
- Train models → Quality metrics: evaluates SSIM
- Train models → Saved checkpoints: saves checkpoints
- Train models → Training previews: writes previews
- Enhance test photos → Training and image utilities: uses test options
- Enhance test photos → Enhancement ResNet: applies enhancement model
- Enhance test photos → Provided checkpoints: loads pretrained models
- Enhance test photos → Saved checkpoints: loads trained models
- Enhance test photos → Enhanced test images: writes enhanced photos