1
Double Takes consists of twenty diptychs whose broad common genre is street photography. On the left side is a “real” photograph that I took, while the right side shows a synthetic image obtained by extracting a description of the former - a so-called “reverse prompt” - using Claude, an LLM (Large Language Model), and then feeding that prompt into an image-generator (Google’s Nano Banana Pro).
Crucially, this prompting technique is purely verbal and non-pictorial, the object of the exercise being to generate an approximation of the “master” image (the original photo) without recourse to any visual suggestion. I requested that the reverse prompts – the description extracted by the LLM – be presented in JSON code, a lightweight, text-based format for representing structured data as name/value pairs and arrays.
The object here is not to achieve the maximum possible coincidence between the two images but to weave a web of repetition and difference with the generation of a synthetic image that teases plausibility and verisimilitude and creates a feeling of uncertainty and uncanniness.
The relationship between words and pictures is both complex and paradoxical : if a picture is worth a thousand words, it is no less true that a word is worth a thousand pictures.