Affective image editing aims to transform an image toward a target emotional state. Existing emotion-based editing methods typically represent emotion as a global condition, which often leads models to produce templated edits, limiting their ability to control fine-grained affective variations. To address these limitations, we propose Affective Continuous Editing Framework (ACE), an object–atmospheric framework for modeling visual emotion. To reduce templated editing, our framework employs an Affective Disentanglement Agent (ADA) to translate abstract affective intent into region-level editing instructions and counterfactual supervision signals. Guided by this data, we de-sign a decoupled regional LoRA module with two independent lightweight branches, Object LoRA (OB-LoRA) and Atmosphere LoRA (AT- LoRA), which respectively learn object and atmospheric editing directions. In addition, we introduce a counterfactual regional suppression loss, encouraging the model to learn disentangled regional features and continuous, fine-grained affective variations. Experiments demon-strate that ACE achieves better identity preservation, region-level controllability, and fine-grained emotional expression than existing emo-tion-conditioned editing methods. Quantitative analyses further confirm that the two learned control directions are monotonic and largely confined to their intended regions. The dataset and code are available at https://github.com/xixixixixxxx/ACE.