description [Jan 19th 2022] Paper released on arXiv.
integration_instructions [Jan 14th 2022] Code released on GitHub.
Neural graphics primitives, parameterized by fully connected neural networks, can be costly to train and evaluate. We reduce this cost with a versatile new input encoding that permits the use of a smaller network without sacrificing quality, thus significantly reducing the number of floating point and memory access operations. A small neural network is augmented by a multiresolution hash table of trainable feature vectors whose values are optimized through stochastic gradient descent. The multiresolution structure allows the network to disambiguate hash collisions, making for a simple architecture that is trivial to parallelize on modern GPUs. We leverage this parallelism by implementing the whole system using fully-fused CUDA kernels with a focus on minimizing wasted bandwidth and compute operations. We achieve a combined speedup of several orders of magnitude, enabling training of high-quality neural graphics primitives in a matter of seconds, and rendering in tens of milliseconds at a resolution of 1920x1080.
A demonstration of the reconstruction quality of different encodings. Each configuration was trained for 11000 steps using our fast NeRF implementation, varying only the input encoding and the neural network size. The number of trainable parameters (neural network weights + encoding parameters) and training time are shown below each image. Our encoding (d) with a similar total number of trainable parameters as the frequency encoding (c) trains over 8 times faster, due to the sparsity of updates to the parameters and smaller neural network. Increasing the number of parameters (e) further improves approximation quality without significantly increasing training time.
Fly-throughs of trained real-world NeRFs. Large, natural 360 scenes (left) as well as complex scenes with many disocclusions and specular surfaces (right) are well supported.
Both models can be rendered in real time and were trained in under 5 minutes from casually captured data: the left one from an iPhone video and the right one from 34 photographs.
Direct visualization of a neural radiance cache, in which the network predicts outgoing radiance at the first non-specular vertex of each pixel's path, and is trained on-line from rays generated by a real-time pathtracer. On the left, we show results using the triangle wave encoding of [Müller et al. 2021]; on the right, the new multiresolution hash encoding allows the network to learn much sharper details, for example in the shadow regions.
Instant Neural Graphics Primitives with a Multiresolution Hash Encoding
Thomas Müller, Alex Evans, Christoph Schied, Alexander Keller