Affective computing has various challenges especially for features extraction. Semantic information and vocal messages contain much emotional information, while extracting affective from features of images, and affective computing for image dataset are regarded as a promised research direction. This paper developed an improved adaptive neuro-fuzzy inference system (ANFIS) for images’ features extraction. Affective value of valence, arousal, and dominance were the proposed system outputs, where the color, morphology, and texture were the inputs. The least-square and k-mean clustering methods were employed as learning algorithms of the system. This improved model for structure and parameter identification of ANFIS were trained and validated. The training errors of the system for the affective values were tested and compared. Data sources grouping and the ANFIS generating processes were included. In the network training process, the number of input variables and fuzzy subset membership function types has been relative to network model under different affective inputs. Finally, well-established training model was used for testing using International Affective Picture System. The resulting network predicted those affective values, which compared to the expected outputs. The results demonstrated the effect of larger deviation of the individual data. In addition, the relationships between training errors, fuzzy sample set, training data set, function type, and the number of membership functions were illustrated. The proposed model showed the effectiveness for image affective extraction modeling with maximum training errors of 14 %.