From Semantic Editing Intentions to Persistent Parametric Controls for AI-Assisted Image Editing

Abstract

AI-powered image editing increasingly allows users to express high-level editing intentions, yet how these intentions are translated into concrete editing actions often remains difficult to understand. We present Atelier, a semantic image editing tool that uses image semantics and user intent to determine which editing controls to surface and how to initialize them. Guided by a formative study with eight image-editing practitioners, Atelier composes context-specific, directly manipulable controls that users can refine rather than committing AI interpretations directly to finished edits. In a within-subjects study with 12 practitioners, participants preferred AI-composed controls, described them as supporting creative exploration, and actively refined AI proposals. However, experiences of control were mixed, and interviews highlighted limited insight into why particular controls and values were proposed. Our findings suggest that making AI interpretations actionable through manipulable controls can support exploration and refinement, but actionability alone does not make those interpretations transparent to all users.