AN INTELLIGENT WEB APPLICATION TO ACCELERATE ART REFERENCE DISCOVERY USING POSE ESTIMATION AND GENERATIVE IMAGE SYNTHESIS

Abstract

Artists depend on reference images to study proportion, lighting, and movement, yet locating a reference that matches a specific visual intention remains slow because mainstream image search matches keywords rather than visual structure. This paper presents Reference Gallery, a web application that lets artists retrieve and generate art references through three channels: a written description, a full-body pose built on a draggable mannequin, and a color chosen from a color wheel. Library poses are indexed with a pose-estimation model and ranked by joint distance, while colors are ranked by distance from each image’s dominant palette. When no stored reference is suitable, a moderated generation service synthesises a new one and adds it to the shared library. An internal ranking study over ten queries produced a mean relevance score of 7.2 out of 10, with pose queries scoring a perfect 10.0 and description and color queries averaging 6.0, indicating that structural channels outperform semantic ones here.