Creating an AI-generated image of myself was a fascinating experience, though it started with an unexpected challenge: describing my own face. I spent far more time than anticipated trying to articulate details that I normally take for granted: my hairstyle, face shape, and the shape of my nose. Since I also wanted a cyberpunk aesthetic, I settled on the following prompt:" A young Chinese woman wearing brown-framed glasses, her black, shoulder-length hair styled without bangs. She has a soft, rounded face with a narrow forehead and a gentle, fleshy nose. The image blends these natural features with a high-tech, cyberpunk aesthetic." This prompt is still not satisfying as though something essential was still missing.
I tested this prompt across three tools: NovelAI, Artlist.io, and Nano Banana. NovelAI leaned too heavily into a Japanese anime style, making me look like a fictional character in manga. Artlist.io consistently ignored my request for "no bangs," stubbornly adding them to every version. Moreover, every image it generated seems to have the generic, average face for a young Chinese woman. Nano Banana produced the most accurate result, yet it still didn't quite capture my true likeness. The part of the image that looked most like me was probably the brown-framed glasses. However, I don't blame the tool; if I struggle to describe myself accurately, I can hardly expect an AI to bridge that gap perfectly. Through this experiment, I’ve grown even more impressed by people who write prompts with precision and skill. Once everyone masters the basic use of AI tools, the quality of the prompt becomes the key factor shaping the outcome.
Wenbin, your systematic approach to testing three different AI tools with the same prompt is excellent, especially because it is this kind of comparative analysis that can reveal important patterns. Your observation that Artlist.io consistently produced "the generic, average face for a young Chinese woman" despite your specific descriptions raises compelling questions about how these systems work.
ReplyDeleteI'm intrigued by your comment that "something essential was still missing" from your prompt. This points to fundamental questions about what aspects of identity can and cannot be translated through algorithmic systems. Your point about Artlist.io "stubbornly" ignoring your "no bangs" request also deserves deeper exploration; why do some systems override user instructions?
You've generated excellent comparative data here. For Assignment 3, I encourage you to connect these observations to our course readings on AI narratives, selfies, and digital identity. The scholarly frameworks we're exploring will help you move from describing what happened to analyzing what it reveals about algorithmic bias, representation, and power.
Looking forward to seeing more in Assignment 3!