This paper introduces PepGLAD, a generative model for full-atom peptide design using geometric latent diffusion. The authors address the challenges of full-atom geometry and variable binding geometry in peptide design, which are not well handled by existing diffusion-based models. PepGLAD employs a variational autoencoder to encode full-atom residues into a fixed-dimensional latent space, allowing for efficient diffusion processes. Additionally, a receptor-specific affine transformation is proposed to convert 3D coordinates into a standard space, enhancing the model's ability to generalize across different binding sites. Experimental results demonstrate that PepGLAD significantly improves diversity, binding affinity, and success rates in sequence-structure co-design tasks, outperforming existing methods. The model also excels in recovering reference structures for binding conformation generation, showcasing its robustness and effectiveness in peptide design.This paper introduces PepGLAD, a generative model for full-atom peptide design using geometric latent diffusion. The authors address the challenges of full-atom geometry and variable binding geometry in peptide design, which are not well handled by existing diffusion-based models. PepGLAD employs a variational autoencoder to encode full-atom residues into a fixed-dimensional latent space, allowing for efficient diffusion processes. Additionally, a receptor-specific affine transformation is proposed to convert 3D coordinates into a standard space, enhancing the model's ability to generalize across different binding sites. Experimental results demonstrate that PepGLAD significantly improves diversity, binding affinity, and success rates in sequence-structure co-design tasks, outperforming existing methods. The model also excels in recovering reference structures for binding conformation generation, showcasing its robustness and effectiveness in peptide design.