Visions is a series of 1000 Generative AI Data Paintings—each probing the elusive borders between fact, fantasy, and the fragile nature of human recollection. Rooted in cognitive science research, the series reimagines our reliance on memory in a time when both the human mind and machine intelligence can seamlessly distort, invent, and reshape what we perceive as real.
At the heart of Visions is an iterative, dual-pipeline AI system that continuously refines and regenerates imagery through a feedback loop. The first pipeline—built on Python, PyTorch, and Transformers—expands images via diffusion models like Flux and Stable Diffusion, segments them using Segment Anything Model (SAM), and contextualizes them with vision-language tools like MiniCPM-V and Florence-2. By alternately outpainting and inpainting, the AI “remembers” and “forgets” objects much like the human mind, offering visual metaphors for the mutable nature of memory. In parallel, a TouchDesigner visualization pipeline renders each iterative stage in real time, revealing the evolving layers of object detection, segmentation masks, and semantic analysis.
Beyond its technological sophistication, Visions serves as a meditation on what it means to see, remember, and believe. Each piece provokes the viewer to question the authenticity of both algorithmic and human recollection—reminding us that our shared sense of reality is often built on precarious, shifting foundations. By merging computational artistry with deeply human questions about truth, emotion, and memory, Visions not only highlights the powerful creative potential of machine intelligence but also invites a broader conversation: in an age when images can be endlessly manufactured, where does genuine perception end and imaginative reconstruction begin?
In a world saturated by images—each vying for our trust—“Visions” explores the entangled nature of human memory, AI-generated imagery, and the biases that distort both. Drawing on the understanding that memory is creative and fragile, this collection pushes us to question how biases in data and algorithms can co-author our recollections. If human memory already reshapes the past in fleeting acts of imagination, AI systems trained on skewed or unrepresentative datasets can further bend what we see and believe.
Each piece in “Visions” is an interplay of manipulated fragments—glitches, reconstructions, and ghostly echoes—underscoring the subtle ways in which AI can perpetuate stereotypes or erase entire narratives. At first glance, these works seem like familiar snapshots of reality; but look closer, and you glimpse how prejudice encoded in a model’s “visual memory” can alter the faces, bodies, or contexts we think we recognize. The collection thus reveals an unsettling truth: in the same way false memories can be implanted in our minds, AI-generated illusions can seed and reinforce new distortions, often leaving us oblivious to their origin.
“Visions” confronts the hyperreal possibilities heralded by deepfakes and synthetic content, showing how the absence of any physical “original” blurs the lines between authenticity and invention. Yet the focus here is not merely the shock of AI-generated worlds, but the quieter, more pervasive effect: the normalizing of biased depictions through repetition. Within this evolving visual ecosystem, memory becomes a collaborative act between mind, machine, and cultural assumption—where illusions can harden into collective myths if unchecked.
By blending techniques like object-recognition, neural-network layering, and generative inpainting, the artworks in “Visions” invite viewers into a space of suspended disbelief. We catch ourselves asking: Which details are algorithmic hallucination, which are relics of personal bias, and which belong to our genuine recollection? This friction between the familiar and the fabricated highlights how easily biases—both human and computational—can rewrite our narratives and alter our emotional truths.
“Visions” stands as a reminder that the power to generate and edit images is also the power to shape belief. Stepping into these AI-fueled visions, we are asked to confront the shared responsibility of creators, technologists, and viewers alike: how do we uphold trust in an era where illusions can be seamlessly woven into our communal memory? This collection offers no simple answers but urges us to remain vigilant. In a time when seeing can mislead, it is in the mindful questioning of each image—especially those generated by partial or prejudiced algorithms—that we preserve the authenticity of what it means to remember, to empathize, and to truly see.
"Visions" implements a dual-pipeline architecture that combines advanced AI image generation and analysis systems with real-time visualization. The project creates an iterative feedback loop where AI-generated content evolves through continuous analysis and regeneration processes.
•Outpainting (Flux): Diffusion-based model expands images to the desired format
•Vision-Language Analysis (MiniCPM-V): Generates textual descriptions of images
•Object Detection (Florence-2): Identifies objects in each image
•Object Filtering & Selection: Applies predefined relevance criteria
•Segmentation Mask Generation (SAM): Creates segmentation masks for selected objects