The paper addresses the limitation of 3D Gaussian Splatting (3DGS) in real-time novel view synthesis, which heavily relies on accurate initialization from Structure-from-Motion (SfM) methods. When the initial point cloud is noisy or randomly initialized, 3DGS often performs poorly. To overcome this, the authors propose RAIN-GS (Relaxing Accurate Initialization Constraint for 3D Gaussian Splatting), a novel optimization strategy that relaxes the need for accurate initialization. RAIN-GS consists of three key components: 1) sparse-large-variance (SLV) random initialization, 2) progressive Gaussian low-pass filtering, and 3) an adaptive bound-expanding split (ABE-Split) algorithm. These components enable 3DGS to prioritize the learning of low-frequency components and encourage Gaussians to move further from their initial positions, effectively relaxing the requirement for accurate initialization. The effectiveness of RAIN-GS is demonstrated through quantitative and qualitative comparisons on multiple datasets, showing that it achieves performance on par with or even better than 3DGS trained with accurate SfM point clouds, even when using randomly initialized point clouds. This opens up new possibilities for applying 3DGS in scenarios where accurate initial point clouds are challenging to obtain.The paper addresses the limitation of 3D Gaussian Splatting (3DGS) in real-time novel view synthesis, which heavily relies on accurate initialization from Structure-from-Motion (SfM) methods. When the initial point cloud is noisy or randomly initialized, 3DGS often performs poorly. To overcome this, the authors propose RAIN-GS (Relaxing Accurate Initialization Constraint for 3D Gaussian Splatting), a novel optimization strategy that relaxes the need for accurate initialization. RAIN-GS consists of three key components: 1) sparse-large-variance (SLV) random initialization, 2) progressive Gaussian low-pass filtering, and 3) an adaptive bound-expanding split (ABE-Split) algorithm. These components enable 3DGS to prioritize the learning of low-frequency components and encourage Gaussians to move further from their initial positions, effectively relaxing the requirement for accurate initialization. The effectiveness of RAIN-GS is demonstrated through quantitative and qualitative comparisons on multiple datasets, showing that it achieves performance on par with or even better than 3DGS trained with accurate SfM point clouds, even when using randomly initialized point clouds. This opens up new possibilities for applying 3DGS in scenarios where accurate initial point clouds are challenging to obtain.