Algorithmic Compositions In Venice | Venice, Italy | Unit 21 | 2020
Awarded Distinction for Design & Distinction for Thesis
This project is an investigation into using Image-to-Image translation algorithms to create architectural design. This process uses hundreds of paired digital images as training for an algorithm to learn information about the chosen input. My project is based in Venice and focuses on the churches in the city as data to inform the algorithm. The significant number of churches provided information for a series of datasets designed to be combined together and create three dimensional buildings from large quantities of two dimensional information. The sets of images aim to explore several different aspects of Venetian churches, from plans to materials, and this method allows the creation of a design that is a mathematical ‘collage’ and visual amalgamation of the history of Venice and its many interesting influences. The information is collated into the design of an art gallery in the city and the project hopes to explore how aspects of machine learning can aid in the creation and design of architecture.
The project began with an analysis of a woodcut print titled ‘View of Venice’, made in the 1500’s by Jacop de’ Barbari. This birds-eye view of the city is in fact a fake and fabricated perspective as at the time it was created nobody would have ever viewed the world from this angle before. On closer inspection of the image you can see several ways in which the artist has distorted the city such as exaggerating the height and scale of certain buildings, distorting buildings and assigning more detail and importance to certain significant places. These elements of the map are designed as such to convey a set of nuanced ideological messages.
This led me to attempt to make my own modern day ‘View of Venice’ by creating a fake satellite map using Pix2Pix – an image to image translation Algorithm. After acquiring the data the maps are then divided into hundreds of separate tiles and paired with their matching tile. The algorithm is then trained on these hundreds of paired images and learns by comparing one to the other. After this training it will create a model which can then be used for me to input my own data afterwards and the algorithm is able to generate a very convincing depiction but it is nonetheless a fake view of the city.
The next stage in my project looked to continue using Image-to-Image translation in a way that could generate architecture. Churches were one of the main buildings highlighted in the View of Venice map and there are over a hundred located in the city meaning there would be a good amount with which to create data sets.
A series of data sets of paired images were made using selected information from 120 churches in Venice. These models feed into each other so the output from the previous can be used as the input for the next. Each is designed to focus on the information we want the algorithm to learn and compare in each instance and makes use of contrasting block colours to highlight each specific element.
This process could in fact, theoretically, be repeated an infinite amount of times with the combination of these steps generating a unique building in every instance. The algorithm works by analysing the probability of placement of pixels within the hundreds of examples it looks at. So the outputs created and combined together give us buildings that are reminiscent of churches and contain many of their key noticeable features but within forms that are manipulated and distorted in many subtle ways.
Originally churches in Venice were typically paid for by the wealthiest member of the local community who would commision local artists to display their work within the building. Perhaps a modern day version of the church is a gallery of sorts and art displayed can be sold to raise money to preserve the existing churches. AI generated art is becoming increasingly popular. An example is this piece of work which was sold at christies in 2018 for 432 and a half thousand dollars and was created with the same type of machine learning I have been using in this project. This new building will aim to offer something to the community that perhaps the existing churches no longer can and at the same time raise money to preserve them from further damage.
A lux level reading was taken of the building indicating which areas within would be suitable for the display of certain types of artwork. This grid pattern was then developed into a gradient between the designated areas and prominent colour from each chosen artwork was used to create a suitable tone.
Using photos I had taken on our trip to Venice I designed a dataset for generating materials. A sample of the picture was taken in the format used for the algorithm and paired with a prominent colour from that material. Once I had the model from this training I could take the colours from the wall design to generate new materials based on all the venetian churches i had visited and photographed.