LLaVA-Emo: Interpretable Affective Image Stylization via Chain-of-Thought Reasoning

Affective Image Stylization (AIS) converts an emotional intent into executable artistic visual styles. Existing methods are often limited to discrete label settings and provide limited interpretability of how target emotions are realized. We propose LLaVA-Emo, an interpretable AIS framework built on multimodal Chain-of-Thought (CoT) reasoning. Our method decouples generation into two structured outputs: <reasoning> provides visual–affective causal explanations grounded in the input image evidence, and <style_prompt> expresses actionable, renderer-ready style instructions that directly condition a frozen diffusion renderer. We constructed a training set by screening ArtEmis’ sentiment interpretations and fine-tune LLaVA-1.5-7B with LoRA, where SFT mainly supervises the structured intermediate <reasoning> (and output format), while the true executability of <style_prompt> is enforced by our DPO stage via render-and-reward feedback. The rendering stage remains training-free, and we further apply DPO for preference optimization to align candidate outputs with both emotion fidelity and instruction executability. Experiments on the EmoEdit inference set demonstrate that LLaVA-Emo improves emotion alignment while providing stronger process interpretability.