DECOMPDIFF is a diffusion model designed for structure-based drug design (SBDD) that incorporates decomposed priors for ligand molecules. The model decomposes a ligand into arms and scaffold, which are responsible for interacting with the target and positioning the arms, respectively. This decomposition allows for more efficient and accurate generation of drug-like molecules by considering the different roles of atoms in the ligand. The model integrates bond diffusion and additional validity guidance during the sampling phase to improve the properties of generated molecules. Extensive experiments on the CrossDocked2020 dataset show that DECOMPDIFF achieves state-of-the-art performance, with an average Vina Dock score of -8.39 and a 24.5% success rate. The model's decomposed priors, bond diffusion, and validity guidance contribute to generating high-affinity molecules with proper molecular properties and conformational stability. The code is available at https://github.com/bytedance/DecompDiff.DECOMPDIFF is a diffusion model designed for structure-based drug design (SBDD) that incorporates decomposed priors for ligand molecules. The model decomposes a ligand into arms and scaffold, which are responsible for interacting with the target and positioning the arms, respectively. This decomposition allows for more efficient and accurate generation of drug-like molecules by considering the different roles of atoms in the ligand. The model integrates bond diffusion and additional validity guidance during the sampling phase to improve the properties of generated molecules. Extensive experiments on the CrossDocked2020 dataset show that DECOMPDIFF achieves state-of-the-art performance, with an average Vina Dock score of -8.39 and a 24.5% success rate. The model's decomposed priors, bond diffusion, and validity guidance contribute to generating high-affinity molecules with proper molecular properties and conformational stability. The code is available at https://github.com/bytedance/DecompDiff.