Product attributes can display the selling points of products, helping users find their desired products in search results. However, product attributes are typically incomplete. In e-commerce, products have multimodal features, including original attributes, images, and texts. How to make full use of the multimodal data to complete the missing attributes is the key challenge. To this end, we propose MPKGAC, a powerful three-stream framework that handles multimodal product data for attribute completion. We build a multimodal product knowledge graph (KG) from the multimodal features, and then convert the attribute completion problem into a multimodal KG completion task. MPKGAC encodes each modality separately, fuses them adaptively, and integrates multimodal decoders for prediction. Experiments show that MPKGAC outperforms the best baseline by 6.2% in Hit@1. MPKGAC is employed to enrich selling points of the women’s clothing industry at Alibaba.com.cn and improves the click-through rate (CTR) by a relative 2.14%.