Automating an Artistic Judgment Without Losing the Artist

The challenge

Composing an image for production depended on artistic judgment that existed only in people's heads. That made the process impossible to scale and inconsistent between operators. Off-the-shelf tools sat at one extreme or the other: purely technical segmentation that ignored composition, or systems needing so much human correction they saved nothing.

Where we came in

Most AI proposals in this space promise to replace the judgment. We designed for the opposite — separate the artistic analysis from the coordinate generation so each stage can be inspected, and keep a person in the loop at the point where taste actually matters. A multi-stage pipeline is also a pipeline you can debug, which single-shot generation never is.

What we did

• Split the problem into artistic analysis and coordinate generation so each stage could be quality-checked independently
• Used established multimodal model APIs rather than training from scratch, reducing technical risk on an already novel problem
• Built an annotation interface where a person reviews, classifies and corrects what the models propose
• Made human and model annotations directly comparable, so model quality could be judged against a human baseline
• Deployed the application to Azure

Where it landed

• A working proof of concept that produces artistic annotations for human review rather than requiring them from scratch
• Model output and human output are captured in the same structure, making quality measurable instead of anecdotal
• Staged architecture means a weak step can be improved without rebuilding the pipeline