Visual attribute as an inter-mediate representation is exploited in many applications due to its semantic understandable and generalization. However, attribute learning need great manual effort for choosing attribute taxonomy and labeling attributes instance, which inevitably introduces human bias and leads to weak discriminant ability of attributes, especially for fine-grained recognition where high discriminant ability is crucial for recognizing subtle distinctness. Motivated by human cognition and the fact of multi-mode distribution of object, the proposed complex attributes try to model the distribution of object explaining various appearance factors and finally forms a distributed representation with better describable and discriminant ability, which will competent for handling high discriminant requirement for fine-grained recognition. Some experiments are conducted on publicly available fine-grained datasets CUB. Results show it has better performance than the handcrafted attributes and also holds simple category discriminant attributes.