How an artist built a self-portrait machine using AI
The process behind an AI-powered self-portrait machine turns a camera, a generative model and a display into a portrait studio. The ITP 30 Show project, associated with NYU’s Interactive Telecommunications Program, asks what happens when an algorithm studies its maker and returns a face for that person to judge.
Rather than presenting AI as a seamless filter, the machine makes its machinery visible. Every image depends on framing, light, training material, written direction and processing speed. Visitors encounter the result and a chain of creative choices, with tension between technological precision and the instability of human likeness.
Starting with a simple camera
At the centre is a camera positioned to reproduce a repeatable view. Its image is captured, prepared and passed to an image-generation system; the result then appears on a nearby screen or printed output. A fixed camera removes many of the accidents involved in taking a phone selfie.
The artist treats this pipeline as an instrument rather than a single app. Choosing the lens, camera height and distance changes the apparent personality of the sitter. Hardware decisions determine whether the interaction feels like a portrait session, a photobooth gag or a thoughtful exchange between artist and machine.
Turning a portrait into data
Before generation, the portrait may be resized, aligned or segmented so the model can interpret the face. Prompts and reference images steer its style, while a fixed random seed can make repeated visits comparable. The system is not inventing from nothing; it recombines learned visual patterns.
Results can range from a close likeness to an unfamiliar mask, and that range is central to the project. The ITP 30 Show’s project archive situates the work among interactive experiments and gives curious viewers the context needed to understand its credits, materials and related program resources.
Teaching the machine a visual language
The visual character comes from a deliberate collection of images rather than an indiscriminate feed. By training or conditioning the system on selected portraits, drawings and textures, the artist gives the machine a vocabulary. The dataset becomes part of the artwork’s authorship, carrying choices about beauty, age, culture and representation.
Written prompts supply another layer. A request for crisp studio lighting, painterly edges or imagined anatomy can push the output beyond the camera image, while repeated testing reveals which terms the model follows reliably. The strongest prompts do not simply describe a style; they create room for productive errors.
Each generated portrait is compared with the source and the artist’s intention. An impressive image can still fail because it flattens the sitter or makes an accident the main event. Refinement happens through looking, editing and trying again, much as in an analogue darkroom.
Making the loop feel responsive
Speed shapes the emotional rhythm. Too much delay breaks the feeling of conversation, while instant output can make the system seem opaque. The artist balances model size, hardware, network access and resolution so the participant can notice the transformation without waiting for a loading screen.
Physical components deserve the same attention as the model. A companion account of component choices in wearable work appears in the show’s wearable hardware notes. In a Melbourne laneway studio, where improvised workspaces are common, that practical mindset turns a clever prototype into something repeatable.
Designing for imperfect rooms
A machine that performs well under perfect studio light can stumble in an ordinary room. Strong Australian daylight, coloured walls and screen glare alter the colours and shapes supplied to the model. Testing across different locations reveals where the portrait pipeline needs calibration and where a strange result may honestly reveal the setting.
Public display adds another test. At Canberra’s National Portrait Gallery, or in a small independent gallery, the camera, screen and labels must guide visitors without turning the encounter into a queue for a novelty photo. Clear instructions and a pause between portraits protect the attention on which the work depends.
Asking harder questions about likeness
Once the system can produce persuasive faces, consent becomes part of the design. The artist must decide who may be recorded, how long images are kept and whether a portrait can be published or sold. Those rules determine the boundary between participation and extraction.
Australian directness can make a useful curatorial tone. Labels might say, “I reckon this looks more like the machine than the person,” inviting viewers to separate resemblance, style and identity without claiming that one image settles the question. The remark acknowledges uncertainty without disappearing into technical jargon.
The resulting work follows familiar Australian routes: limited editions through independent galleries, direct commissions and online stores with prices in Australian dollars. The artist’s name and the model’s material history need to travel with the object, not remain buried in production notes.
Sharing the work with a wider public
The online format lets the project reach viewers beyond New York, where ITP’s history and media culture often dominate discussion. Australian audiences can approach the self-portrait machine through questions shared with local creative communities: how prototypes are funded, how experimental media gains an audience and how authorship is credited.
Documentation is therefore part of the exhibit. Project pages, author credits, event listings and RSVP information help remote visitors understand both the finished interaction and the collaboration behind it. Showing failed outputs, wiring diagrams or test images can reveal the process, making the machine less like magic and more like an argument made visible.
Explore the ITP 30 Show online, read the project materials and attend the associated event if one is listed. Follow the artist’s credits and related program resources to see experimental art, design and code meet in a working piece rather than a polished demonstration.
Make a portrait, study the face on the screen and treat it as the beginning of a conversation about creative authorship, data and the limits of machine-made likeness.