Knowledge graphs (KGs) often suffer from missing links, hindering downstream tasks that rely on their completeness. In order to complete the KGs, existing knowledge graph completion (KGC) methods primarily leverage semantic or structural information within the KG. However, relying solely on this information often leads to suboptimal prediction performance. To overcome this limitation, we propose KG-diffusion, a novel KGC method designed to generate missing links and thereby enhance KG completion. KG-diffusion employs an encoder-decoder architecture operating in a continuous vector space. The encoder was used to learn KG embeddings, while Gaussian noise is incrementally introduced into the entity embeddings. The decoder’s objective is to reconstruct the original embeddings from these noise representations. Recognizing the importance of conditioning in the generative process, we introduce an aggregator module that enhances conditioning by integrating information from neighboring entities. This enriched conditioning process guides the decoder towards more accurate generation. Extensive evaluations on two publicly available benchmark datasets demonstrate that KG-diffusion achieves state-of-the-art performance in KGC tasks.