EmoStyle: Affective Conditioning of Style-Specialist Experts for Emotional Image Generation
Year: 2026
Type: conference-paper
Abstract: Affective image editing must change perceived emotion without disturbing the identity, scene, and structure that make an image recognizable. Existing editors can follow textual instructions, but an emotion request rarely specifies which visual evidence should change or which content should remain fixed. We introduce IECLA, a framework that separates source-emotion evidence from target-specific editing control. IECLA learns signed regional contributions from raw counterfactual supervision and turns eligible evidence into constrained local instructions for existing editors. On a locked teacher-defined test set, IECLA attains an importance Spearman correlation of 0.5885 and an R2 of 0.4005. On the independent Human320 benchmark, it reaches importance and signed Spearman correlations of 0.5001 and 0.4525. Across 5,440 fixed editing tasks and five editors, IECLA routes 1,869 eligible cross-emotion cases, retains the native output otherwise, and raises end-to-end target-success point estimates by 5.3%--17.9% relative to the corresponding native pipelines. These results indicate that separating where source emotion is expressed from how a target emotion is rendered improves affective control without modifying an editor's weights or sampling procedure.
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7168293136