Michela Coslovich: Your artistic research focuses on the complexity of computer-generated imagery (CGI): how do they relate to contemporary digital culture?

Sheung Yiu: That is a big question and deserves a much more in-depth answer, but I will try to keep it brief. First, let me clarify, when I say CGI, I don’t just mean fancy visual effects in games and movies. I am also referring to a new genre of images created or sustained by a myriad of computational techniques. That includes metadata-embedded smartphone photos, to AI-powered beautifying filters that apply effects on your self-facing camera by default even without you using any filters (in the case of Tiktok earlier this year), to Deepfakes and digital objects created by 3D photogrammetry.

CGI and contemporary digital culture are inexplicably intertwined, and they affect and inform the development of each other. For one, I think the most obvious characteristic of CGI that relates to contemporary digital culture is its networked nature. Images are grouped on social media platforms by albums, stock photography sites by categories, and computer vision datasets by labels and tagging. The network culture not only affects how images are disseminated and how they create meanings. An image is rarely understood alone, especially in digital culture. Instead, it is always understood together with other images, connected through sensor data and human labels, no matter how visually unrelated they may seem on the surface. Images are connected to other images, but not only that, they are connected to a digital ecology consisting of computational models, human users, and machines. A selfie unsuspectingly uploaded on Flickr is scrapped into a training dataset for facial recognition. The algorithm developed is sold to a state agency, which is then used by law enforcement to identify political activists, or by a soldier controlling an aerial drone. The ripple effect of an image is convoluted, and spread across time and space beyond the moment of capture and upload. In fact, I have moved from using the term CGI to ‘algorithmic image system’ to denote the hyper-connectedness of computational images.

Another aspect concerns scales, both in the horizontal and vertical sense. By horizontal scale, I mean the abundance of data and the surplus of online images create what-we-now-call big data. The emergence of big data provides the necessary materials for statistical predictions and pattern recognition. Facebook most famously exploits users’ profile data and their online interactions to predict future behaviors and sell targeted content to unsuspecting individuals — surveillance capitalism. Similar logic is being applied to computer vision, where through analyzing large amounts of images, computer scientists establish a statistical correlation between a particular pixel pattern with a semantic conclusion. Digital culture and visual culture are both experiencing a statistical turn. 

By vertical scale, I mean the hyper-connectedness of images and the omnipresence of cameras make it possible to capture the same objects from different distances in different positions with different resolutions. For some popular visual subjects, we have captured so many images and data that now we can observe it from the pictorial level all the way down to the pixel level or even deep layers (deep in the neural network sense of deep learning). The many scales in which we capture and observe the same subject has implications on how much more information we can extract from an image. The Google computer vision project NeRF in the Wild uses a special type of neural network to interpolate 3D structure of famous monuments from all the tourist photos taken from different vantage points. In remote sensing, a mathematical correlation between pixel value and tree size established through ground-level observation helps scientists interpret satellite images of trees taken from space, revealing information that is otherwise lost due to resolution limits. Even though individual trees are invisible in a satellite image in which one pixel represents a 10m x 10m area of land, scientists can recover, or ‘resurrect’, a tree by applying the same statistical model established through comparing ground-level observation with satellite data. In other words, this new way of trans-scalar seeing allows us to almost limitlessly extrapolate and interpolate information encoded on a digital image, converting between human observations and machine calculations and traversing between human scale and planetary scale.