When searching for an image, sometimes you have a picture but want something slightly different, like a similar scene but with a different color. Traditional systems struggle because they don't explicitly understand the requested change.
This work introduces FactEmbed, which breaks down the query into separate factors representing different visual attributes. It first uses these factors to imagine what the target image should look like, then uses that imagination to build a retrieval embedding. This separates understanding the query from creating the final search representation.
FactEmbed outperforms previous approaches across standard benchmarks, meaning it more accurately finds the intended image. This improves user experience in applications like e-commerce or photo libraries where users modify reference images with text.