Neural radiance fields (NeRFs) enable high-quality novel view synthesis, but
their prohibitively high computational complexity limits deployability,
especially on resource-constrained platforms. To enable practical usage of
NeRFs, quality tuning is essential to reduce computational complexity, akin to
adjustable graphics settings in video games. However while existing solutions
strive for efficiency, they use one-size-fits-all architectures regardless of
scene complexity, although the same architecture may be unnecessarily large for
simple scenes but insufficient for complex ones. Thus as NeRFs become more
widely used for 3D visualization, there is a need to dynamically optimize the
neural network component of NeRFs to achieve a balance between computational
complexity and specific targets for synthesis quality. Addressing this gap, we
introduce NAS-NeRF: a generative neural architecture search strategy uniquely
tailored to generate NeRF architectures on a per-scene basis by optimizing the
trade-off between complexity and performance, while adhering to constraints on
computational budget and minimum synthesis quality. Our experiments on the
Blender synthetic dataset show the proposed NAS-NeRF can generate architectures
up to 5.74$\times$ smaller, with 4.19$\times$ fewer FLOPs, and 1.93$\times$
faster on a GPU than baseline NeRFs, without suffering a drop in SSIM.
Furthermore, we illustrate that NAS-NeRF can also achieve architectures up to
23$\times$ smaller, 22$\times$ fewer FLOPs, and 4.7$\times$ faster than
baseline NeRFs with only a 5.3\% average SSIM drop. The source code for our
work is also made publicly available at
this https URL

通过为每个场景基于复杂度和性能的权衡优化神经网络组件，我们提出了 NAS-NeRF：一种唯一定制化的生成式神经架构搜索策略，以在计算预算和最小合成质量的约束条件下动态优化 NeRF 架构。实验结果表明，NAS-NeRF 相比于基准 NeRF 可生成的架构规模减小多达 5.74 倍，FLOPs 减少 4.19 倍，并且在 GPU 上速度提升 1.93 倍，而不会降低结构相似性。同时，NAS-NeRF 还可以实现比基准 NeRF 小 23 倍的架构，FLOPs 减少 22 倍，并且只有 5.3% 的平均结构相似性下降。