The paper "Score Distillation via Reparametrized DDIM" addresses the issue of over-smoothing and unrealistic outputs in 3D shape generation using Score Distillation Sampling (SDS). The authors analyze that SDS can be understood as a high-variance version of Denoising Diffusion Implicit Models (DDIM), where SDS introduces noise independently at each step, while DDIM infers noise conditionally from previous predictions. This difference leads to excessive variance and poor quality in 3D generation. To improve this, the authors propose Score Distillation via Inversion (SDI), which replaces the random noise sampling in SDS with prompt-conditioned DDIM inversion. This modification aligns the SDS update rule more closely with DDIM, reducing variance and improving the quality of 3D generations. Experimental results show that SDI achieves similar or better 3D generation quality compared to other state-of-the-art methods, without requiring additional neural networks or multi-view supervision. The key contributions include a theoretical analysis of the relationship between SDS and DDIM, and the introduction of SDI as a simple and effective solution to the over-smoothing issue in 3D generation.The paper "Score Distillation via Reparametrized DDIM" addresses the issue of over-smoothing and unrealistic outputs in 3D shape generation using Score Distillation Sampling (SDS). The authors analyze that SDS can be understood as a high-variance version of Denoising Diffusion Implicit Models (DDIM), where SDS introduces noise independently at each step, while DDIM infers noise conditionally from previous predictions. This difference leads to excessive variance and poor quality in 3D generation. To improve this, the authors propose Score Distillation via Inversion (SDI), which replaces the random noise sampling in SDS with prompt-conditioned DDIM inversion. This modification aligns the SDS update rule more closely with DDIM, reducing variance and improving the quality of 3D generations. Experimental results show that SDI achieves similar or better 3D generation quality compared to other state-of-the-art methods, without requiring additional neural networks or multi-view supervision. The key contributions include a theoretical analysis of the relationship between SDS and DDIM, and the introduction of SDI as a simple and effective solution to the over-smoothing issue in 3D generation.