GPT Image Generation Test: Good at Recoloring, Weak at Composition

snikolov · x · 2026-08-08

A designer shared an experiment on X involving iterative collage generation with GPT. The workflow involved prompting the model to generate different colorings for patches of source material, which were then assembled into a new collage.

The author observed that GPT handles color editing well because it is a relatively objective task. However, when asked to rearrange the patches into a new, interesting composition, the model struggled, as subjective layout and aesthetic restructuring remain a weak point.

Original post →

More from Models

Models channel →