The　single　label　of　an　affective　image　cannot　well　reflect　the　emotions　underneath,　thus　converting　the　single　label　into　compound　affective　labels　better　reveals　the　complex　emotions　of　the　image.　However,　building　new　datasets　on　human　annotation　has　high　labor　costs,　and　the　results　are　quite　subjective.　To　address　this　issue,　our　paper　proposes　a　model　for　constructing　compound　affective　labels　from　single-label　datasets　with　a　CNN-based　model.　First,　we　enhance　the　ability　of　image　sentiment　classification　by　applying　a　new　classifier　loss.　Second,　we　adopt　the　knowledge　distillation　model　for　retaining　more　sentiment　information　in　image　labels.　Third,　we　integrate　the　attention　mechanism　within　the　knowledge　distillation　to　transfer　the　attention　map　to　the　student　network　for　improving　its　performance.　Finally,　we　attain　the　compound　affective　labels　from　the　label　probability　distribution　of　the　distilled　model.　The　generated　image　datasets　with　compound　labels　can　be　applied　in　various　fields.　They　can　serve　for　psychological　analysis　and　evaluation,　and　provide　richer　affective　references　in　art　design,　interactive　media,　advertising　products,　etc.