In this paper, we present GEARS, a generative approach to quickly editing a real image in ways that target specific objects or image regions. We consider the scenario where a user has an initial image and wants to specify fine-grained “difference” prompts (e.g., “the carpet is more worn”). When these edits are intended to improve image retrieval, it is vital to avoid adding new features or diffusing the impact of the edit in making small variations throughout the image. Therefore, we introduce an approach that translates natural language edits into masks that limit the scope of the edits, allowing a user to selectively and iteratively update the image. We show example results and insights from our tests in using these edits for image retrieval in the domain of hotel recognition.