Ask most AI image models for a portrait and you get an instant result — no visible process, no record of how it arrived, a different picture entirely if you ask again. Every Face Project, a solo-authored paper by Vinod Rajpurohit posted to arXiv on November 6, is built deliberately against that grain: a live, working portrait studio that makes the sitter watch the drawing happen, and that produces a document you can replay afterward.
How a sitting actually works
A visitor picks one of eight fictional artists, each written with its own city, medium, and tempo — distinct aesthetic personas rather than a single uniform “style” toggle. They hold still for 8 to 12 seconds while a still image is captured. With consent, that still is sent to a hosted image model, which returns a private “study” — an intermediate result the sitter never actually sees. A compiler then turns that hidden study into a single drawing document, which plays out live over 44 seconds as the sitter watches it render.
That compressed, visible unfolding is the whole point: rather than a picture simply appearing, the sitter experiences something that behaves like watching an artist actually draw — a beginning, a progression, a specific duration — even though what’s actually happening underneath is a generative model’s output being revealed through a compiled playback, not a brush moving in real time. Because the result is a document rather than a flat image, the paper describes it as replayable: the same sitting can be watched again later, which is a meaningfully different object than a saved JPEG.
An unreviewed preprint, by design worth taking seriously
The paper carries no journal reference or conference acceptance listed — an unreviewed preprint, not a peer-reviewed publication — but it ships with something a lot of HCI preprints don’t: a live system, at everyfaceproject.com, that a reader can actually try rather than taking the paper’s word for. For a project whose entire argument is about the experience of sitting for it, that’s the right kind of evidence to offer.