Mip-NeRF 360 is an extension of mip-NeRF, a variant of Neural Radiance Fields (NeRF) designed to address sampling and aliasing issues. This extension targets unbounded scenes, where the camera can point in any direction and content can exist at any distance. The key contributions of Mip-NeRF 360 include: 1. **Scene Parameterization**: A non-linear scene parameterization that maps 3D coordinates to a contracted space, ensuring that distant points are distributed according to disparity rather than distance. This helps in balancing the detail and scale of nearby and distant objects. 2. **Online Distillation**: An efficient training strategy that uses a small "proposal MLP" to predict volumetric density and guide the training of a larger "NeRF MLP." This approach reduces the training time while maintaining high-quality rendering. 3. **Regularization**: A distortion-based regularizer designed for mip-NeRF ray intervals to prevent artifacts such as "floaters" and "background collapse." The model significantly improves upon previous methods, reducing mean squared error by 57% compared to mip-NeRF and producing realistic synthesized views and detailed depth maps for complex, unbounded real-world scenes. The paper also includes a comprehensive evaluation on a new dataset and an ablation study to demonstrate the effectiveness of each component.Mip-NeRF 360 is an extension of mip-NeRF, a variant of Neural Radiance Fields (NeRF) designed to address sampling and aliasing issues. This extension targets unbounded scenes, where the camera can point in any direction and content can exist at any distance. The key contributions of Mip-NeRF 360 include: 1. **Scene Parameterization**: A non-linear scene parameterization that maps 3D coordinates to a contracted space, ensuring that distant points are distributed according to disparity rather than distance. This helps in balancing the detail and scale of nearby and distant objects. 2. **Online Distillation**: An efficient training strategy that uses a small "proposal MLP" to predict volumetric density and guide the training of a larger "NeRF MLP." This approach reduces the training time while maintaining high-quality rendering. 3. **Regularization**: A distortion-based regularizer designed for mip-NeRF ray intervals to prevent artifacts such as "floaters" and "background collapse." The model significantly improves upon previous methods, reducing mean squared error by 57% compared to mip-NeRF and producing realistic synthesized views and detailed depth maps for complex, unbounded real-world scenes. The paper also includes a comprehensive evaluation on a new dataset and an ablation study to demonstrate the effectiveness of each component.