The paper “Seeding diversity into AI Art”, written by Mr Marvin Zammit, Dr Antonios Liapis, and Prof Georgios Yannakakis from the Institute of Digital Games of the University of Malta has been awarded Best Student Paper at the 13th International Conference on Computational Creativity. The award is granted to the best paper whose first author is also a student.
ICCC 2022 took place in Bolzano, Italy, between the 27th June and the 1stof July, and it was held in person for the first time after the COVID pandemic. It is the only scientific conference that focuses on all aspects of computational processes which creates music, narrative, images, games, humour, or any other product which can be deemed creative in nature.
This year, the conference hosted 55 papers, 5 workshops and tutorials, and a Doctoral Consortium Program for over 20 Doctoral candidates from all over the world.
The winning paper was presented by Mr Marvin Zammit during the conference and presented a novel method of using genetic algorithms in conjunction with text-to-image generation based on deep learning to produce more diverse images for the same text prompt. In 2021 OpenAI revealed a deep learning model called DALL-E which could generate an image from any given text prompt with an unprecedented accuracy. The model itself was not released, but a component of it, CLIP, which can determine how closely a given text matches an image, was made available to the general public. This was quickly put to use by the scientific communities in a method, dubbed VQGAN+CLIP, which could generate images from text which were not quite as accurate as DALL-E, but still visually impressive. Artists, makers, and enthusiasts have developed many applications, artistic installations, twitterbots and more that produce art based on prompts using CLIP.
In their paper, Zammit et al. note that this method tends to generate images which are similar to the human eye for the same text prompt. This is due to the fact that the process generates every single image as a distinct process, with no awareness of prior creations. The novel algorithm proposed by Zammit et al. interrupts the image generation process and introduces an algorithm inspired by natural evolution (which has created massive biodiversity over millenia) in order to inject diversity into the images being composed. This diversity-driven evolutionary algorithm is similar to what happens in, for example, locomotion in animals: there are several forms across different species, e.g. running, galloping, crawling, swimming, flying, etc. Not all of these are in direct competition to each other so all forms survive, given that they excel within their respective niche. Evolutionary algorithms were chosen in this case because they all deal with a population simultaneously, and hence can operate on all images being created at the same time, thereby endowing the generation process with visibility of multiple images.
The Institute of Digital Games is the centre for research and education in game design, game analysis, and game technology at the University of Malta. Our work is at the forefront of innovative games research. We explore games and play, uncovering new playful and generative possibilities in game design and technology. We delve into everything games can teach us about ourselves. Our multidisciplinary academic team spans computer science, literature, game design, philosophy, media studies, and social sciences.
Read our latest report: IDG Annual Report - Academic Year 2020/21
Sample images generated for different text prompts; from left to right: “a lonely house in the woods”, “a pyramid made of ice”, “cosmic love and attention”, “fire in the sky”, and “artificial intelligence”.
Read the awarded paper: https://arxiv.org/abs/2205.00804