Prior Art
alan
I was watching a crime procedural tonight. I was thinking that I would’ve benn a pretty good sketch artist in another life. Then I was hit by a dull inevitability: Forensic Sketch Artists are goners. They’ll be replaced by mugshot GPT.
I’m fully on the side of the artists here, and I don’t presume to know the artistry involved - but I suspect it has a lot to do with the dialogue. Asking the right questions to get into a flow, timing the questions with parts of the face being generated by the artist. I like to imagine it’s similar to improv.
That said, LLMs could knock this out of the park. Even accounting for all the complexities of language and culture and bias, there’s a straightforward reason why this is a job for GPT: the feedback loop.
- Ask questions about the subject of the sketch (and I/O that GPT already does pretty well)
- “Draw” and iterate on the sketch, hallucinating through variations until an optimal feature is surfaced.
- Disseminate the sketch.
- Find the subject
- Wait
- Feed the photo of the subject into the model along with the sketch, dialogue, and other data.
Whether or not the sketch is accurate, the model will improve if the subject is found and their face is treated as training data. It’s dark, but I could see a pipeline from description to facial recognition and automatic identification.