Users often make purchases with specific intentions, such as decorating a garden or preparing birthday gifts, and their purchases are typically centered around these intentions. Since these underlying intentions are usually not directly observable, leveraging them effectively for Sequential Recommendation (SR) becomes challenging. Existing studies primarily learn intent representations in the latent space by introducing sequence-level data augmentation. However, this augmentation may introduce noise and thus significantly impact items purchased consecutively under the same latent intention. Moreover, the fine-grained representations within user intents, such as information about other items that share the same intent representation as the target item is also overlooked. This information is crucial for the model’s ability to learn user intentions accurately, yet it remains unobserved. To address this issue, we propose the Latent Guided Diffusion model for Sequential Recommendation (LGD4Rec) to learn users’ intent representations. Specifically, we first cluster to find the intention most similar to the recently purchased item, then use latent guided diffusion to capture the representation of this shared intention. Finally, we generate the embedding of the item that the user is likely to purchase under the same intention. This generated item, which considers user intentions, has a certain degree of interpretability, enhancing the robustness of SR. Extensive experiments conducted on four datasets confirm the effectiveness of LGD4Rec and its superior performance compared to established baselines.