Make-A-Shape is a large-scale 3D generative model trained on over 10 million diverse 3D shapes. It addresses the challenges of training large 3D generative models by introducing a wavelet-tree representation, which is compact, expressive, and efficient for training. The wavelet-tree representation encodes 3D shapes using wavelet decomposition, retaining both coarse and detailed information. The model is trained using a diffusion process, with techniques like subband coefficient filtering and packing to exploit coefficient relations and enable efficient training. The framework supports conditional generation from various modalities, including single/multi-view images, point clouds, and voxels. Extensive experiments demonstrate the model's superior performance in unconditional and conditional generation tasks, achieving high-quality results with fast inference times.Make-A-Shape is a large-scale 3D generative model trained on over 10 million diverse 3D shapes. It addresses the challenges of training large 3D generative models by introducing a wavelet-tree representation, which is compact, expressive, and efficient for training. The wavelet-tree representation encodes 3D shapes using wavelet decomposition, retaining both coarse and detailed information. The model is trained using a diffusion process, with techniques like subband coefficient filtering and packing to exploit coefficient relations and enable efficient training. The framework supports conditional generation from various modalities, including single/multi-view images, point clouds, and voxels. Extensive experiments demonstrate the model's superior performance in unconditional and conditional generation tasks, achieving high-quality results with fast inference times.