This page in Swedish

Deep Art Challenge – generating images using AI

Deep Art Challenge.

Join the Deep Art Challenge and learn how to train an AI model to generate image content resembling the paintings of your favourite artist. The challenge is open to all students and does not require previous experience with AI development or programming. It also contains more advanced topics for those with a background in computer science, or those eager to dig deeper into AI development.

Participating is easy

The challenge is divided into three different parts and anyone can participate, no matter the field of study or previous technical skills. The first part requires no programming knowledge. Results are generated by following step-by-step instructions. Experimenting with parameter settings is what will alter the results. The second part does not require knowledge in programming either. It is all about working with the input data and creating your own dataset for training the machine learning model. The third part is typically directed at computer science students who want a programming challenge. You don't have to complete all three parts to participate in the challenge.

What you will learn

  • Train a deep learning model.
  • Interpret training results.
  • The importance of training data in AI development.
  • Run python code using Jupyter Notebook.
  • Work in the popular data science platform Kaggle.

We dive directly into deep learning, that is multi-layered neural networks.

Computer vision is one of the most common applications of deep learning. It is all around you. Take your mobile phone and turn the front camera on. A bounding box tracking your face will most likely appear on screen. Object detection in images is commonly solved with a deep learning algorithm. It is quite straightforward how it works. We feed the image into a deep neural network. It is called deep because it consists of multiple connected layers. Each layer detects patterns or features of a certain level, and subsequent layers can combine features of previous layers to detect higher level features. At the last step the algoritm makes a prediction of what class the detected objects belong to, in this case a human face. What features and classes can be detected depends on what features and classes were present in the dataset we trained the model with. If you are new to machine learning you will soon realise that the quality of the training dataset is very important.

So what happens if we run this kind of algorithm in reverse? Would it then create images? We are now entering the domain of generative machine learning, and more specifically Deep Convolutional Generative Adversarial Networks (DCGAN). A DCGAN takes noise as input, and upsamples and re-distributes the input at multiple layers in a deep neural network to produce an image resembling the style of the images in the dataset used when training the model. Let's try this! More detailed explanations of the technique is provided in the instructions on Kaggle.

Your goal is to train a DCGAN model to generate low-resolution images resembling paintings of the American artist Bob Ross.

1. Create a free account on the data science platform Kaggle:

2. Open a copy of the project:

3. This should open a Jupyter Notebook with instructions on how to move forward.

4. When you are happy with your results, download your five favourites among the generated images and email them to This is an email address, also include a link to your notebook and a short description of what tweaks you did to hyperparameters, code, input data etc.

Your goal is to use the DCGAN introduced in the first part, and replace the Bob Ross image dataset with your own dataset, to generate images resembling the style of your own favourite artist.

1. Collect images of your favourite artist, or a set of images of the same concept or style.

2. Open a copy of the project:

3. This should open a Jupyter Notebook with instructions on how to move forward.

4. When you are happy with your results, download five examples of your collected images, and your five favourites among the generated images and email them to This is an email address, also include a link to your notebook and a short description of your image dataset, including the number of collected images.

Now you are on your own. We challenge you to go beyond the DCGAN architecture and find a way to generate high-resolution images resembling the paintings of Bob Ross. You could start, for example, by looking up BigGAN, or a style transfer technique like CycleGAN.

Send five examples of your generated images to This is an email address, and include a link to your code and a short description of your approach.

We will have an art display at campus, as well as publish your results in a gallery on our webpage. It will be updated continously as we receive your contributions. We will keep the challenge open over the summer and reward the best contributions in the autumn. Contributions will be judged by looks, effort, and innovativeness.

If you have technical issues or other questions, please send an email to This is an email address.