instant-ngp：将神经渲染加速几个数量级

Instant Neural Graphics Primitives with a Multiresolution Hash Encoding

https://arxiv.org/abs/2201.05989

https://github.com/NVlabs/instant-ngp

https://nvlabs.github.io/instant-ngp/

6.9k stars

Authors: Thomas Müller, Alex Evans, Christoph Schied, Alexander Keller

摘要：由完全连接的神经网络参数化的神经图形原语，训练和评估成本高昂。我们通过一种通用的新输入编码来降低成本，该编码允许使用较小的网络而不牺牲质量，从而显著减少浮点和内存访问操作的数量：一个小型神经网络通过一个可训练特征向量的多分辨率哈希表来扩充，该哈希表的值通过随机梯度下降进行优化。多分辨率结构允许网络消除哈希冲突的歧义，从而形成一个简单的体系结构，在现代GPU上进行并行化非常简单。我们利用这种并行性，使用完全融合的CUDA内核实现整个系统，重点是最小化浪费的带宽和计算操作。我们实现了几个数量级的组合加速，能够在几秒钟内训练出高质量的神经图形原语，并以1920×1080的分辨率在数十毫秒内渲染。△ 较少的

Submitted 4 May, 2022; v1 submitted 16 January, 2022; originally announced January 2022.

Comments: To appear in ACM Transactions on Graphics (SIGGRAPH 2022). 15 pages, 13 figures, 3 tables

Journal ref: ACM Trans. Graph. 41, 4, Article 102 (July 2022), 15 pages

@article{mueller2022instant,

    author = {Thomas M\"uller and Alex Evans and Christoph Schied and Alexander Keller},

    title = {Instant Neural Graphics Primitives with a Multiresolution Hash Encoding},

    journal = {ACM Trans. Graph.},

    issue_date = {July 2022},

    volume = {41},

    number = {4},

    month = jul,

    year = {2022},

    pages = {102:1--102:15},

    articleno = {102},

    numpages = {15},

    url = {https://doi.org/10.1145/3528223.3530127},

    doi = {10.1145/3528223.3530127},

    publisher = {ACM},

    address = {New York, NY, USA},

}