The field of 3D shape generation has recently seen a surge in interest, with researchers exploring various techniques to create high-dimensional 3D shapes. The rise of various 3D shape representation methods has brought about the challenge of integrating these forms into neural networks for training purposes. Lately, the shape generation community has relied on encoding these representations into low-dimensional latent spaces using trained encoder models and synthesizing new shapes using latent generative models. However, this approach has limitations, such as the high computational power needed to train an auto-encoder. In contrast, this work proposes an innovative integration of wavelet decomposition and Singular Value Decomposition (SVD) as a neural network-free approach to capturing the essence of 3D shapes. The proposed method efficiently embeds 3D shapes into a robust, differentiable implicit representation without needing a prior training phase. The key advantage of this representation is its suitability for scenarios with limited data, where conventional generative models may struggle. By leveraging a score-matching generative model, the proposed pipeline demonstrates the ability to create high-quality 3D shapes that match the performance of leading methods but with less reliance on extensive datasets that typically constrain the creation of intricate 3D geometries. Moreover, it achieves superior MMD, COV, and JS scores in industry-relevant cases such as Rims and for the ShapeNet dataset. Remarkably, it also reduces the GPU power needed by [60% ~ 90%] compared to different models.     «

The field of 3D shape generation has recently seen a surge in interest, with researchers exploring various techniques to create high-dimensional 3D shapes. The rise of various 3D shape representation methods has brought about the challenge of integrating these forms into neural networks for training purposes. Lately, the shape generation community has relied on encoding these representations into low-dimensional latent spaces using trained encoder models and synthesizing new shapes using latent...     »