Emotion Evoked Image Generation with Personalized Emotional Cues

Qing Lin¹, Mengmi Zhang¹; ¹Nanyang Technological University

Despite rapid progress in generative visual models, emotion evoked image generation remains under explored. Images elicit emotional responses through their semantics, context, and structural composition, making this problem important for applications such as mental health support, product design, and creative practice. The central challenge is to modify a given image so that it evokes a target emotion while preserving the semantic content and structural coherence of the original scene. We study emotion evoked image generation and propose a diffusion based framework that explicitly models emotionally salient visual cues while incorporating personalized human feedback. The method identifies regions and elements that contribute most strongly to perceived emotion and selectively transforms them to guide emotional responses. Through simple evaluative signals, the framework adapts to individual affective preferences, enabling the same image to evoke the same target emotion in a manner consistent with different observers. We evaluate the proposed approach through systematic human perceptual experiments assessing emotional clarity, semantic preservation, and overall preference. Across multiple emotions and diverse images, human observers consistently judge the generated results to evoke clearer intended emotions while better preserving the original scenes compared to existing baselines. These results demonstrate that explicit and personalized modeling of emotional cues substantially improves emotion evoked image generation and highlight emotional understanding as a core capability for future generative models.

Acknowledgements: This research is supported by the National Research Foundation, Singapore under its NRFF award NRF-NRFF15-2023-0001 and Mengmi Zhang's Startup Grant from Nanyang Technological University, Singapore.

Thank You to Our AVS 2026 Sponsors