Search papers, labs, and topics across Lattice.
This paper introduces AffectDelta, an innovative image editing framework that utilizes eight-dimensional emotion distributions to facilitate nuanced emotion-driven modifications in images while maintaining compositional integrity. By employing a frozen Emotion Distribution Predictor to assess the source image's emotional state and a source-aware diffusion backbone for context-sensitive transformations, AffectDelta significantly enhances the precision of emotional transitions compared to traditional single-label methods. Experimental results show that AffectDelta achieves superior affective alignment and content preservation, validated through comprehensive quantitative and qualitative evaluations against six baseline models.
Emotion-driven image editing just got a major upgrade鈥擜ffectDelta captures complex emotional transitions with unprecedented precision using eight-dimensional distributions.
Emotion-driven image editing aims to evoke a specified target emotion by modifying emotion-relevant visual cues in a source image, while preserving the overall composition and semantic-structural coherence of the original scene. Existing scene-level editors typically specify the target with a single emotion category and often learn visual transformations from operation-level text instructions. A category collapses a mixed affective endpoint into one dominant label, while language cannot precisely quantify how coexisting emotions should increase, decrease, or remain stable. We introduce AffectDelta, a source-aware editor that treats editing as a transition between eight-dimensional emotion distributions. A frozen Emotion Distribution Predictor estimates the source state, and the signed source-to-target difference encodes the direction and magnitude of the requested transition. Within AffectDelta, an internal transition encoder and a source-aware diffusion backbone jointly translate this signal into context-dependent semantic and appearance changes. To train this formulation, we construct AffectPair-249K, comprising 248,841 source-target pairs with predicted eight-dimensional distributions and spanning both cross-category and within-category transitions. Experiments against six baselines, combining quantitative evaluation with qualitative comparisons, demonstrate improved affective alignment and content preservation, while ablations validate our design choices. Code and dataset will be made publicly available upon acceptance.
