AbsGS: Recovering Fine Details for 3D Gaussian Splatting The paper introduces AbsGS, a method to address the over-reconstruction issue in 3D Gaussian Splatting (3D-GS). The original adaptive density control strategy in 3D-GS suffers from gradient collision, leading to blurry images in regions with high-frequency details. AbsGS proposes homodirectional view-space positional gradient as a criterion for densification, which avoids gradient collision and enables the splitting of large Gaussians in over-reconstructed regions. The method improves rendering quality and reduces memory consumption compared to 3D-GS. Experiments on various datasets show that AbsGS achieves better results in terms of PSNR, SSIM, and LPIPS, while maintaining similar or lower memory usage. The method is easy to implement and can be integrated into existing Gaussian Splatting-based methods. The project is open-sourced.AbsGS: Recovering Fine Details for 3D Gaussian Splatting The paper introduces AbsGS, a method to address the over-reconstruction issue in 3D Gaussian Splatting (3D-GS). The original adaptive density control strategy in 3D-GS suffers from gradient collision, leading to blurry images in regions with high-frequency details. AbsGS proposes homodirectional view-space positional gradient as a criterion for densification, which avoids gradient collision and enables the splitting of large Gaussians in over-reconstructed regions. The method improves rendering quality and reduces memory consumption compared to 3D-GS. Experiments on various datasets show that AbsGS achieves better results in terms of PSNR, SSIM, and LPIPS, while maintaining similar or lower memory usage. The method is easy to implement and can be integrated into existing Gaussian Splatting-based methods. The project is open-sourced.