What Can I Edit? Open-Ended Strategy Discovery and the Emotion Editability Landscape

Traditional affective image manipulation (AIM) methods rely on bounded strategy spaces, such as predefined factor taxonomies or knowledge libraries. These approaches often fail to capture context-grounded, image-specific strategies that humans find meaningful. This paper asks: can we move beyond these fixed templates by reframing the task from 'how should I edit?' to 'what can I edit?'

The authors introduce EmoScope, a multi-agent framework that treats emotional image editing as an open-ended discovery task. The system first performs emotion-conditioned affordance reasoning to identify what elements within a specific image can be modified to evoke a target emotion. It then uses a semantic hierarchy—consisting of anchors (elements to preserve), variables (elements to modify), and context—to navigate the trade-off between content consistency and emotional expressiveness. The framework employs two iterative loops: a planning loop that generates and verifies the editing strategy, and an editing loop that executes and refines the image generation. Because the plan is expressed as affordances, it also provides an interactive surface for users to refine the emotional strategy at a high level.

In a large-scale human evaluation involving 4,693 responses across eight emotion categories, participants preferred EmoScope over competitive baselines by an average of 88.1%. Attribution analysis confirms that EmoScope generates adaptive, context-aware strategies rather than applying uniform templates. The authors also demonstrate that existing classifier-based metrics (like Emo-A and Emo-S) suffer from 'blind spots' toward non-stereotypical, context-grounded edits that humans prefer. Finally, the study provides an empirical map of the 'content-emotion preference-affinity landscape,' showing that the effectiveness of emotional editing is not uniform but varies significantly depending on the specific image content and target emotion.