We present Gen3C, a generative video model with precise Camera Control and temporal 3D Consistency. Prior video models already generate realistic videos, but they tend to leverage little 3D information, leading to inconsistencies, such as objects popping in and out of existence. Camera control, if implemented at all, is imprecise, because camera parameters are mere inputs to the neural network which must then infer how the video depends on the camera. In contrast, Gen3C is guided by a 3D cache: point clouds obtained by predicting the pixel-wise depth of seed images or previously generated frames. When generating the next frames, Gen3C is conditioned on the 2D renderings of the 3D cache with the new camera trajectory provided by the user. Crucially, this means that Gen3C neither has to remember what it previously generated nor does it have to infer the image structure from the camera pose. The model, instead, can focus all its generative power on previously unobserved regions, as well as advancing the scene state to the next frame. Our results demonstrate more precise camera control than prior work, as well as state-of-the-art results in sparse-view novel view synthesis, even in challenging settings such as driving scenes and monocular dynamic video. Results are best viewed in videos. Check out our webpage!1

Neural fields are a highly effective representation across visual computing. This work observes that fitting these fields is greatly improved by incorporating spatial stochasticity during training, and that this simple technique can replace or even outperform custom-designed hierarchies and frequency-space constructions. The approach is formalized as implicitly operating on a blurred version of the field, evaluated in-expectation by sampling with Gaussian-distributed offsets. Querying the blurred field during optimization greatly improves convergence and robustness, akin to the role of preconditioners in numerical linear algebra. This implicit, sampling-based perspective fits naturally into the neural field paradigm, comes at no additional cost, and is extremely simple to implement. We describe the basic theory of this technique, including details such as handling boundary conditions, and extending to a spatially-varying blur. Experiments demonstrate this approach on representations including coordinate MLPs, neural hashgrids, triplanes, and more, across tasks including surface reconstruction and radiance fields. In settings where custom-designed hierarchies have already been developed, stochastic preconditioning nearly matches or improves their performance with a simple and unified approach; in settings without existing hierarchies it provides an immediate boost to quality and robustness.

Neural graphics primitives are faster and achieve higher quality when their neural networks are augmented by spatial data structures that hold trainable features arranged in a grid. However, existing feature grids either come with a large memory footprint (dense or factorized grids, trees, and hash tables) or slow performance (index learning and vector quantization). In this paper, we show that a hash table with learned probes has neither disadvantage, resulting in a favorable combination of size and speed. Inference is faster than unprobed hash tables at equal quality while training is only 1.2–2.6 × slower, significantly outperforming prior index learning approaches. We arrive at this formulation by casting all feature grids into a common framework: they each correspond to a lookup function that indexes into a table of feature vectors. In this framework, the lookup functions of existing data structures can be combined by simple arithmetic combinations of their indices, resulting in Pareto optimal compression and speed.

Neural radiance fields achieve unprecedented quality for novel view synthesis, but their volumetric formulation remains expensive, requiring a huge number of samples to render high-resolution images. Volumetric encodings are essential to represent fuzzy geometry such as foliage and hair, and they are well-suited for stochastic optimization. Yet, many scenes ultimately consist largely of solid surfaces which can be accurately rendered by a single sample per pixel. Based on this insight, we propose a neural radiance formulation that smoothly transitions between volumetric- and surface-based rendering, greatly accelerating rendering speed and even improving visual fidelity. Our method constructs an explicit mesh envelope which spatially bounds a neural volumetric representation. In solid regions, the envelope nearly converges to a surface and can often be rendered with a single sample. To this end, we generalize the NeuS [Wang et al. 2021] formulation with a learned spatially-varying kernel size which encodes the spread of the density, fitting a wide kernel to volume-like regions and a tight kernel to surface-like regions. We then extract an explicit mesh of a narrow band around the surface, with width determined by the kernel size, and fine-tune the radiance field within this band. At inference time, we cast rays against the mesh and evaluate the radiance field only within the enclosed region, greatly reducing the number of samples required. Experiments show that our approach enables efficient rendering at very high fidelity. We also demonstrate that the extracted envelope enables downstream applications such as animation and simulation.

Volumetric representations are popular in inverse rendering because they have a simple parameterization, are smoothly varying, and transparently handle topology changes. However, incorporating the full volumetric transport of light is costly and challenging, often leading practitioners to implement simplified models, such as purely emissive and absorbing volumes with "baked" lighting. One such challenge is the efficient estimation of the gradients of the volume's appearance with respect to its scattering and absorption parameters. We show that the straightforward approach---differentiating a volumetric free-flight sampler---can lead to biased and high-variance gradients, hindering optimization. Instead, we propose using a new sampling strategy: differential ratio tracking , which is unbiased, yields low-variance gradients, and runs in linear time. Differential ratio tracking combines ratio tracking and reservoir sampling to estimate gradients by sampling distances proportional to the unweighted transmittance rather than the usual extinction-weighted transmittance. In addition, we observe local minima when optimizing scattering parameters to reproduce dense volumes or surfaces. We show that these local minima can be overcome by bootstrapping the optimization from nonphysical emissive volumes that are easily optimized.

Physically based differentiable rendering has recently evolved into a powerful tool for solving inverse problems involving light. Methods in this area perform a differentiable simulation of the physical process of light transport and scattering to estimate partial derivatives relating scene parameters to pixels in the rendered image. Together with gradient-based optimization, such algorithms have interesting applications in diverse disciplines, e.g., to improve the reconstruction of 3D scenes, while accounting for interreflection and transparency, or to design meta-materials with specified optical properties. The most versatile differentiable rendering algorithms rely on reverse-mode differentiation to compute all requested derivatives at once, enabling optimization of scene descriptions with millions of free parameters. However, a severe limitation of the reverse-mode approach is that it requires a detailed transcript of the computation that is subsequently replayed to back-propagate derivatives to the scene parameters. The transcript of typical renderings is extremely large, exceeding the available system memory by many orders of magnitude, hence current methods are limited to simple scenes rendered at low resolutions and sample counts. We introduce radiative backpropagation , a fundamentally different approach to differentiable rendering that does not require a transcript, greatly improving its scalability and efficiency. Our main insight is that reverse-mode propagation through a rendering algorithm can be interpreted as the solution of a continuous transport problem involving the partial derivative of radiance with respect to the optimization objective. This quantity is "emitted" by sensors, "scattered" by the scene, and eventually "received" by objects with differentiable parameters. Differentiable rendering then decomposes into two separate primal and adjoint simulation steps that scale to complex scenes rendered at high resolutions. We also investigated biased variants of this algorithm and find that they considerably improve both runtime and convergence speed. We showcase an efficient GPU implementation of radiative backpropagation and compare its performance and the quality of its gradients to prior work.

Modern rendering systems are confronted with a dauntingly large and growing set of requirements: in their pursuit of realism, physically based techniques must increasingly account for intricate properties of light, such as its spectral composition or polarization. To reduce prohibitive rendering times, vectorized renderers exploit coherence via instruction-level parallelism on CPUs and GPUs. Differentiable rendering algorithms propagate derivatives through a simulation to optimize an objective function, e.g., to reconstruct a scene from reference images. Catering to such diverse use cases is challenging and has led to numerous purpose-built systems---partly, because retrofitting features of this complexity onto an existing renderer involves an error-prone and infeasibly intrusive transformation of elementary data structures, interfaces between components, and their implementations (in other words, everything). We propose Mitsuba 2, a versatile renderer that is intrinsically retargetable to various applications including the ones listed above. Mitsuba 2 is implemented in modern C++ and leverages template metaprogramming to replace types and instrument the control flow of components such as BSDFs, volumes, emitters, and rendering algorithms. At compile time, it automatically transforms arithmetic, data structures, and function dispatch, turning generic algorithms into a variety of efficient implementations without the tedium of manual redesign. Possible transformations include changing the representation of color, generating a "wide" renderer that operates on bundles of light paths, just-in-time compilation to create computational kernels that run on the GPU, and forward/reverse-mode automatic differentiation. Transformations can be chained, which further enriches the space of algorithms derived from a single generic implementation. We demonstrate the effectiveness and simplicity of our approach on several applications that would be very challenging to create without assistance: a rendering algorithm based on coherent MCMC exploration, a caustic design method for gradient-index optics, and a technique for reconstructing heterogeneous media in the presence of multiple scattering.