Original article was published by Dariusz Gross #DATAsculptor on Artificial Intelligence on Medium

His sculptures were figurative, religious in style and subject. The main material he used was wood. As part of the project, I aim to have his work materialised with the help of an industrial robot, which will have the ability to sculpt in my father’s style with the same tools.

2020 I received a scholarship grant from the Minister of Culture. It concerns preparatory works for building a database. To this end, the website was created. Please do visit — on this page you are able see the components of this database, i.e. 3D scans, as well as spherical panoramas, all of which are available on this page. You can see where my father’s works are displayed — which, due to the specificity of his work, is very important. It was also important for me to document the time and place it all began — it seemed to me that it may be important for training the model in the future.

2020 #WimieOJCA

All scanned models are accessible through the browser. You can see in 3D what my father’s works look like. Additionally, there is a newsletter on the website. If you are interested in the topic of data sculpting, you are welcome to subscribe. You can receive the latest news concerning not only data sculpting, but also GANs’ and machine learning’s use in art.

The most important part of the project is the database. It consists of several thousand photos of sculptures. These photos are then further processed. Each element of the photo is very important during the model training. You will see it on the examples of sculptures being generated that the background is important, the lighting is important, and everything from whether the photo is coloured and how it is formatted plays a key part in the process. I prepare these photos in such a way that they do not lose their proportions — this is important for modelling. I normalise the lighting, I ‘augment’ the dataset by processing the photos, including symmetrical reflections, and as I mentioned earlier — the change of lighting, change of colour. I use different methods to make the final effect as close as possible to my father’s style.


Generation of sculptures

I use various GANs to generate sculptures, including StyleGAN2(NVidia). I have prepared a database and the model, looking at the database consisting of photos of my father’s figures, generates new figures which are similar in style to my father’s figures. This video shows how the GAN learns from the database. This training process requires a lot of time and outputs pictures . The database also includes abstract sculptures made more recently, which expand this database and allow the works being created now to, although being in my father’s style, evoke a new quality. This is what GAN training looks like, and this is a video example where all the generated images have been arranged into a film. They show the process of how generation takes place from abstraction to figurative representation.


see the whole process (code + more experiments) in this video //english subtitles

thank you for reading, you may also enjoy some other articles:

All images created by the author.