The paper introduces Mip-NeRF, a multiscale representation for neural radiance fields (NeRF) that addresses the issue of aliasing and blurring in NeRF renderings. NeRF, a popular method for learning and rendering 3D scenes, samples rays per pixel, which can lead to aliasing and blurring when the scene is observed at different resolutions. To mitigate this, Mip-NeRF extends NeRF to represent the scene at a continuous scale by rendering anti-aliased conical frustums instead of rays. This approach reduces aliasing artifacts and improves the representation of fine details while being 7% faster and half the size of NeRF. On a challenging multiscale dataset, Mip-NeRF reduces average error rates by 17% compared to NeRF and by 60% compared to a multiscale variant of NeRF. Mip-NeRF also matches the accuracy of a brute-force supersampled NeRF while being 22 times faster. The key contributions include the use of conical frustum tracing and integrated positional encoding (IPE) to explicitly model the size and shape of each frustum, allowing for a single multiscale MLP that combines the functionality of NeRF's "coarse" and "fine" MLPs.The paper introduces Mip-NeRF, a multiscale representation for neural radiance fields (NeRF) that addresses the issue of aliasing and blurring in NeRF renderings. NeRF, a popular method for learning and rendering 3D scenes, samples rays per pixel, which can lead to aliasing and blurring when the scene is observed at different resolutions. To mitigate this, Mip-NeRF extends NeRF to represent the scene at a continuous scale by rendering anti-aliased conical frustums instead of rays. This approach reduces aliasing artifacts and improves the representation of fine details while being 7% faster and half the size of NeRF. On a challenging multiscale dataset, Mip-NeRF reduces average error rates by 17% compared to NeRF and by 60% compared to a multiscale variant of NeRF. Mip-NeRF also matches the accuracy of a brute-force supersampled NeRF while being 22 times faster. The key contributions include the use of conical frustum tracing and integrated positional encoding (IPE) to explicitly model the size and shape of each frustum, allowing for a single multiscale MLP that combines the functionality of NeRF's "coarse" and "fine" MLPs.