Trim 3D Gaussian Splatting (TrimGS) is introduced to reconstruct accurate 3D geometry from images. Unlike previous methods that focus on strong geometric regularization, TrimGS uses Gaussian trimming to selectively remove inaccurate geometry while preserving accurate structures. It analyzes the contributions of individual 3D Gaussians and proposes a contribution-based trimming strategy to remove redundant or inaccurate Gaussians. Experimental and theoretical analyses show that small Gaussian scales are crucial for representing intricate details. TrimGS maintains relatively small Gaussian scales and is compatible with existing geometry regularization strategies. When combined with 3DGS and 2DGS, TrimGS consistently yields more accurate geometry and higher perceptual quality. The method introduces a contribution-based trimming strategy, scale control for better contribution evaluation, and normal regularization to enhance geometry. It also demonstrates that small Gaussian scales improve rendering quality and reduce storage consumption. TrimGS achieves significant improvements in mesh and point cloud evaluation, as well as in rendering quality, particularly in high-frequency regions. The method is compatible with both 3DGS and 2DGS and shows better performance in terms of perceptual quality and storage efficiency. Theoretical analysis shows that small Gaussian scales lead to larger gradients, improving convergence. TrimGS is effective in reducing noise and improving geometry accuracy, making it a valuable technique for 3D Gaussian reconstruction.Trim 3D Gaussian Splatting (TrimGS) is introduced to reconstruct accurate 3D geometry from images. Unlike previous methods that focus on strong geometric regularization, TrimGS uses Gaussian trimming to selectively remove inaccurate geometry while preserving accurate structures. It analyzes the contributions of individual 3D Gaussians and proposes a contribution-based trimming strategy to remove redundant or inaccurate Gaussians. Experimental and theoretical analyses show that small Gaussian scales are crucial for representing intricate details. TrimGS maintains relatively small Gaussian scales and is compatible with existing geometry regularization strategies. When combined with 3DGS and 2DGS, TrimGS consistently yields more accurate geometry and higher perceptual quality. The method introduces a contribution-based trimming strategy, scale control for better contribution evaluation, and normal regularization to enhance geometry. It also demonstrates that small Gaussian scales improve rendering quality and reduce storage consumption. TrimGS achieves significant improvements in mesh and point cloud evaluation, as well as in rendering quality, particularly in high-frequency regions. The method is compatible with both 3DGS and 2DGS and shows better performance in terms of perceptual quality and storage efficiency. Theoretical analysis shows that small Gaussian scales lead to larger gradients, improving convergence. TrimGS is effective in reducing noise and improving geometry accuracy, making it a valuable technique for 3D Gaussian reconstruction.