AI Art with Training Image Experiment of Runway ML
How to create an animated video of synthetically generated StyleGAN landscapes and cities using RunwayML
Collective Memory of a Conflict’ is one which allows us to visualize the development of the Syrian Civil War (2011 — ) through images and videos collected by our contributors. The imagery is event-specific, as it focuses on the incidents which changed the prospects of the war.
The preliminary stage of our project was to collect the necessary images with certain tags related to the topic through its different dimensions. We created a dataset with 3,066 images related to the topic “Syria’s destroyed buildings”. Finally, we practiced the design experiments using RunwayML training section.
But first… Let me introduce you RunwayML!
A Machine Learning Medium: RunwayML
RunwayML is a machine learning medium which brings an accessable visual interface to generate creative models for non-coders and for creators. For example, you can create a new video using Green Screen tool in RunwayML. In addition, you can detect objects, transform content, generate media, chain models, train an image generation model and train an object detection model.
For our project, Collective Memory of a Conflict, we used “Train an image generation model”.
Train an Image Generation Model
In the machine learning part there is “train” section. In this section, we selected “image” because we were experimenting on data set with only images.
In the image generative section there are training options. To use StyleGAN, you should click on advanced. For our project we used StyleGAN2 for generating photorealistic images. In contrast to StyleGAN, StyleGAN2 provides improvements in image quality and variety over StyleGAN.
On the right, there is advanced options in which you can adjust training steps. The more the training steps, the better the results you will get. Actually, it is recommended to use a dataset with 5000 images for a good and neat result. When running models in Runway’s cloud infrastructure, you’ll be running them on fast GPU-enabled computers, hence the estimated training time changes. To start, you should is to click “start training” and then RunwayML will do the job. You do not have to keep RunwayML running while training is on, once it starts you can close RunwayML or shut down your computer. In any case, RunwayML will continue to train and once it is finished, you will be informed via e-mail.
In our project we trained 3066 images with 3000 steps and the training took approximately 3 hours.
And here is an example of a final results from one of our experiments with StyleGAN2:
You may not be a developer, coder, or software engineer, but do not let that stop you from being creative and utilize the latest AI experiments!
