Generative models for structure-based drug design (SBDD) have shown promising results inrecent years. Existing works mainly focus onhow to generate molecules with higher bindingaffinity, ignoring the feasibility prerequisites andresulting in false positives. We conduct thoroughstudies on key factors of ill-conformational problems when applying autoregressive methods anddiffusion to SBDD, including mode collapse andincoherent continuous-discrete space. In this paper, we introduce MolCRAFT, the first SBDDmodel that operates in the continuous parameterspace, together with a novel denoising samplingstrategy. Empirical results show that our modelconsistently outperforms other strong baselines inbinding affinity while achieving more stable 3Dstructure, demonstrating our ability to accuratelymodel interatomic interactions. To our best knowledge, MolCRAFT is the first to achieve reference-level Vina Scores (-6.59 kcal/mol), outperformingother strong baselines by a wide margin (-0.84kcal/mol).