Go to ICLR 2025 Conference homepageGeneralizable Monocular 3D Human Rendering via Direct Gaussian Attribute Diffusion
This paper leverages 3D Gaussian Splatting to tackle the challenging task of generating novel views of humans from given single-view images. Existing methods typically adopt an indirect supervision manner, i.e., splat-based rasterization for differentiable rendering. However, the intricate coupling of various 3D Gaussian attributes complicates precise error backpropagation during optimization, often resulting in convergence to local optima. In contrast, we propose a novel direct paradigm to train a conditional diffusion model directly supervised by proxy-ground-truth 3D Gaussian attributes. Specifically, we propose a two-stage construction process to derive consistent and smoothly distributed proxy-ground-truth 3D Gaussian attributes. Subsequently, we train a point-based conditional diffusion model customized to learn the data distribution of these proxy attributes. The resulting diffusion model can generate the 3D Gaussian attributes for the input single-view image, which are further rendered into novel views. Extensive experimental results showcase the significant performance advancement of our method over state-of-the-art approaches. Source code will be made publicly available.