Deviating from Statistics偏离统计
This is an ongoing series of still images developed as part of my practice-based PhD research. Visually, the work revolves around my own lived experiences of identity and feelings. In terms of the creative thinking behind it, the project is more concerned with how I negotiate with and push against the homogenised aesthetics brought about by generative AI, especially those models trained through the large-scale extraction of online data, which often continue to produce visual tendencies that drift toward an averaged norm. Throughout this series, I have been continuously thinking about how more private, marginal, and difficult-to-standardise feelings can still be preserved when dominant visual languages are increasingly shaped by the scale of data and algorithmic logic. The work has slowly grown through an ongoing process of deviation, selection, rejection, and reconstruction. At the same time, it also records how I gradually embossing my own visual grammar back into the image, layer by layer.
The impulse behind this series began with a very strong feeling that emerged during my use of AI. At the moment, almost all technology companies seem to be moving toward highly similar goals in model training. They continue to pursue image outputs that approach a kind of “perfection”, or, more precisely, a visual standard convincing enough to pass as real. This goal-oriented logic often leads AI-generated images to become overly “correct” in a physical and visual sense. Many of these images are polished, complete, and almost impossible to fault, yet in doing so they also lose something that I personally understand as artistic quality, by which I mean an immaterial space capable of holding emotion and more complex forms of feeling. When creating with these AI models, I often experience a very familiar kind of pressure, a pressure produced by the ongoing friction between minority experience and mainstream aesthetics. The averaged visual language drawn from large-scale datasets often comes into conflict with the sense of loneliness, strangeness, and marginal emotion that I am trying to reach.
Throughout the making process, I do not treat AI as a simple tool that merely “produces images” for me. For me, it functions more like a probe that continuously exposes visual habits and aesthetic inertia. Many times, my real work does not lie in selecting images that I am satisfied with, but in rejecting large numbers of them. I generate, overturn, and regenerate repeatedly, deliberately observing how AI consistently tends to translate emotion into certain standardised visual outcomes, and then consciously finding ways to move around them. A very important part of this project lies precisely in these images that I continuously deny and discard. They allow me to see more clearly what this so-called “average” actually looks like.
On a technical level, AI feels much closer to an artistic material than a tool. To me, it is more like a painter’s pigment, or the clay in the hands of a ceramic maker. I treat AI outputs as raw material, and then through LoRA, image selection, and the processes of recomposition, overlay, and manual intervention in Photoshop, I press my own visual language back into the image. Many of the abstract textures that appear throughout the work are part of my own personal visual vocabulary. A large number of the organic textures, structural forms, and dense atmospheric layers are gradually rebuilt during post-production, piece by piece. For me, this is fundamentally a process of reclaiming authority over the image. Within an increasingly averaged and standardised visual environment, the question of how to still make space for personal experience and individual feeling remains unavoidable when creating with AI, especially with closed-source systems. Those moments of deviation, refusal, repeated revision, and unwillingness to compromise often become the very place where the work truly begins to come into being.