VastGaussian is a novel method for high-quality and real-time large scene reconstruction and rendering based on 3D Gaussian Splatting (3DGS). It addresses the limitations of existing NeRF-based methods in visual quality and rendering speed for large scenes. The key contributions of VastGaussian include: 1. **Progressive Data Partitioning**: The method divides a large scene into multiple cells, optimizing each cell independently before merging them into a complete scene. This approach reduces memory requirements and optimization time, especially when using multiple GPUs. 2. **Decoupled Appearance Modeling**: A novel module is introduced to capture appearance variations in the training images. By attaching appearance embeddings to the rendered image and feeding them into a CNN, the model learns transformation maps to adjust the rendered images, ensuring consistent appearance across different views. 3. **Seamless Merging**: After optimizing individual cells, they are merged seamlessly without noticeable border artifacts. This is achieved by deleting 3D Gaussians outside the original region and ensuring common cameras between adjacent cells. 4. **Performance and Efficiency**: VastGaussian outperforms state-of-the-art methods in terms of visual quality, rendering speed, and training efficiency. It achieves real-time rendering at 1080p resolution with significantly shorter training times and lower video memory consumption. The method is evaluated on several large scene benchmarks, demonstrating superior performance in reconstruction quality and rendering speed. The paper also includes a detailed analysis of the method's components, ablation studies, and supplementary material, providing insights into the CNN architecture, training details, and visualization of transformation maps.VastGaussian is a novel method for high-quality and real-time large scene reconstruction and rendering based on 3D Gaussian Splatting (3DGS). It addresses the limitations of existing NeRF-based methods in visual quality and rendering speed for large scenes. The key contributions of VastGaussian include: 1. **Progressive Data Partitioning**: The method divides a large scene into multiple cells, optimizing each cell independently before merging them into a complete scene. This approach reduces memory requirements and optimization time, especially when using multiple GPUs. 2. **Decoupled Appearance Modeling**: A novel module is introduced to capture appearance variations in the training images. By attaching appearance embeddings to the rendered image and feeding them into a CNN, the model learns transformation maps to adjust the rendered images, ensuring consistent appearance across different views. 3. **Seamless Merging**: After optimizing individual cells, they are merged seamlessly without noticeable border artifacts. This is achieved by deleting 3D Gaussians outside the original region and ensuring common cameras between adjacent cells. 4. **Performance and Efficiency**: VastGaussian outperforms state-of-the-art methods in terms of visual quality, rendering speed, and training efficiency. It achieves real-time rendering at 1080p resolution with significantly shorter training times and lower video memory consumption. The method is evaluated on several large scene benchmarks, demonstrating superior performance in reconstruction quality and rendering speed. The paper also includes a detailed analysis of the method's components, ablation studies, and supplementary material, providing insights into the CNN architecture, training details, and visualization of transformation maps.