Neural Film

Emulate any color style if you have the training data :)

Usage

• install required packages
• Prepare your dataset
  • An example (recommended): export original images from Lightroom as inputs, apply a specific camera profile and then export as labels
  • IMPORTANT: the corresponding input and label must have the same filename
• Place the dataset according to the following structure

  neuralfilm/
  │
  ├─ data/
  │    ├─ <your_dataset>/ - put your training data here
  │    │    ├─ input/
  │    │    └─ label/
  │    └─ <your_dataset_split>/ - will appear once you run split_image.py
  └── ...
  

• Split images into patches

  python split_image.py -d data/<your_dataset>
  

• Modify data_dir in config.json: "data_dir": "data/<your_dataset_split>",
• Train python train.py -c config.json
• Apply the learned filter to any images

  python apply.py -i <input_image>.jpg -r saved/models/neuralfilm/<time>/model_best.pth
  

  • Recommend that <input_image>.jpg is obtained in the same way as the training input (such as in the above example: export from Lightroom without editing)

Acknowledgements

Created from PyTorch Template Project by Victor Huang