This paper introduces an implicit field decoder, called IM-NET, for learning generative models of 3D shapes. IM-NET is trained to assign values to points in 3D space, enabling the extraction of a shape as an iso-surface. The decoder uses a binary classifier to determine whether a point is inside or outside a shape. By replacing conventional decoders with IM-NET, the authors demonstrate superior results in tasks such as generative shape modeling, interpolation, and single-view 3D reconstruction, particularly in terms of visual quality. IM-NET is integrated into various frameworks, including autoencoders (AEs), variational autoencoders (VAEs), and generative adversarial networks (GANs), leading to IM-AEs and IM-GANs. These models show improved performance in shape representation learning, 2D and 3D shape generation, shape interpolation, and single-view 3D reconstruction. The authors also compare their results with state-of-the-art methods, showing that IM-NET produces higher quality shapes with smoother and more coherent surfaces. The method is evaluated using various metrics, including LFD, which is found to be a better visual similarity metric for 3D shapes. The results show that IM-NET outperforms other methods in terms of visual quality and surface quality. The method is also effective in 2D shape generation and interpolation, and in single-view 3D reconstruction. The authors conclude that IM-NET provides a simple and effective approach for learning shape boundaries and has potential for future applications in shape segmentation and correspondence.This paper introduces an implicit field decoder, called IM-NET, for learning generative models of 3D shapes. IM-NET is trained to assign values to points in 3D space, enabling the extraction of a shape as an iso-surface. The decoder uses a binary classifier to determine whether a point is inside or outside a shape. By replacing conventional decoders with IM-NET, the authors demonstrate superior results in tasks such as generative shape modeling, interpolation, and single-view 3D reconstruction, particularly in terms of visual quality. IM-NET is integrated into various frameworks, including autoencoders (AEs), variational autoencoders (VAEs), and generative adversarial networks (GANs), leading to IM-AEs and IM-GANs. These models show improved performance in shape representation learning, 2D and 3D shape generation, shape interpolation, and single-view 3D reconstruction. The authors also compare their results with state-of-the-art methods, showing that IM-NET produces higher quality shapes with smoother and more coherent surfaces. The method is evaluated using various metrics, including LFD, which is found to be a better visual similarity metric for 3D shapes. The results show that IM-NET outperforms other methods in terms of visual quality and surface quality. The method is also effective in 2D shape generation and interpolation, and in single-view 3D reconstruction. The authors conclude that IM-NET provides a simple and effective approach for learning shape boundaries and has potential for future applications in shape segmentation and correspondence.