Denis Zavadski: [email protected]

pip install -r requirements.txt

Use the graphical interface to upload style and contet images. The images will be automatically resized to $128px\times128px$. The model takes only 400 MB, so it should work on most low-end devices. This is what an hypothetical final product would look like.

# Be sure to be in the custom environment created in Setup
cd demo/
streamlit run demo-app.py -- --model_path=checkpoints/RGB_Latent32_best_model.pth.tar

We also provide a Jupyter Notebook that will easily allow the user to generate plots such as those can be found in our Results section. The way to go to have a better comparison between multiple outputs.

implicit-UNet-AdaIN/main.ipynb is the jupyter notebook we used to import the two datasets and to set up and train the network. There is a lot of stuff here, but you find easy access to the datasets and let's you experiment with the training procedure.

Refere to the README inside the `implicit-UNet-AdaIN/" folder for further instructions.

NOTE: a Huggingface token is needed to run the training / download the dataset (this is not necessary to run the forward pass of the network). Create a .env file in the main directory and add HUGGING_FACE_TOKEN='YOUR TOKEN'.

├── demo/
│   ├── checkpoints/   <--------- model checkpoint
│   ├── images/    <-------------- some content and style images for testing
│   ├── demo-app.py
│   └── demo.ipynb
└── implicit-UNet-AdaIN/
    ├── implicit_scarsity_experiments/    <--- plots and tables of loss values
    ├── samples/    <------------------------- testing samples
    ├── samples_organized/    <--------------- testing samples in a more human readable format
    └── main.ipynb

1. Introduction
2. Results

For a detailed insight on the project, please refer to Report.pdf.

Colourization of grayscale images has been an area of significant interest in computer vision, with applications ranging from photo restoration to artistic expression. We propose an approach to colorize grayscale images in various artistic styles using a U-Net architecture enhanced with Adaptive Instance Normalization (AdaIN) layers. U-Net, known for its effectiveness in semantic segmentation tasks, provides an ideal framework for our colorization task due to its ability to capture spatial dependencies while preserving fine details. By incorporating AdaIN layers into the U-Net architecture, we introduce the capability to adaptively transfer artistic styles (here we use style to refer to color choices) from reference images to grayscale inputs. AdaIN enables the decoupling of content and style in feature representations, allowing us to leverage the content information from grayscale images while infusing them with the stylistic characteristics extracted from reference coloured images. This style-guided colourization approach opens up new possibilities for artistic expression, allowing users to apply various painting styles, from impressionism to surrealism, to grayscale inputs.

Here are some examples for some random test images and random paintings. The first row always shows the colourization obtained by passing the ground truth image to the network, like in training phase. That is not a realistic scenario, since in practice the coloured image is not available. On the other hand, it provides a good baseline for comparison.

We encourage the users to experimnt themselves with any image of their choice through the demo-app or the demo.ipynb or otherwise with images from the imagenet-1k and wikiart datasets via main.ipynb.