SIGGRAPH '23 Conf Proc: Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Proceedings

We introduce a geometric multigrid method for solving linear systems arising from variational problems on surfaces in geometry processing, Gravo MG. Our scheme uses point clouds as a reduced representation of the levels of the multigrid hierarchy to achieve a fast hierarchy construction and to extend the applicability of the method from triangle meshes to other surface representations like point clouds, nonmanifold meshes, and polygonal meshes. To build the prolongation operators, we associate each point of the hierarchy to a triangle constructed from points in the next coarser level. We obtain well-shaped candidate triangles by computing graph Voronoi diagrams centered around the coarse points and determining neighboring Voronoi cells. Our selection of triangles ensures that the connections of each point to points at adjacent coarser and finer levels are balanced in the tangential directions. As a result, we obtain sparse prolongation matrices with three entries per row and fast convergence of the solver. Code is available at https://graphics.tudelft.nl/gravo_mg.

We propose a general convex optimization problem for computing regularized geodesic distances. We show that under mild conditions on the regularizer the problem is well posed. We propose three different regularizers and provide analytical solutions in special cases, as well as corresponding efficient optimization algorithms. Additionally, we show how to generalize the approach to the all pairs case by formulating the problem on the product manifold, which leads to symmetric distances. Our regularized distances compare favorably to existing methods, in terms of robustness and ease of calibration.

We propose new methods for combining NDFs in microfacet theory, enabling a wider range of surface statistics. The new BSDFs that follow allow for independent adjustment of appearance at grazing angles, and can’t be represented by linear blends of single-NDF BSDFs. We derive importance sampling for a symmetric operator that blends NDFs uniformly, and introduce a new asymmetric operator that supports NDF variation with elevation. We also extend Smith’s model to support piecewise-constant NDF and material variations with elevation, and demonstrate accuracy via Monte Carlo simulations.

Node graph systems are used ubiquitously for material design in computer graphics. They allow the use of visual programming to achieve desired effects without writing code. As high-level design tools they provide convenience and flexibility, but mastering the creation of node graphs usually requires professional training. We propose an algorithm capable of generating multiple node graphs from different types of prompts, significantly lowering the bar for users to explore a specific design space. Previous work [Guerrero et al. 2022] was limited to unconditional generation of random node graphs, making the generation of an envisioned material challenging. We propose a multi-modal node graph generation neural architecture for high-quality procedural material synthesis which can be conditioned on different inputs (text or image prompts), using a CLIP-based encoder. We also create a substantially augmented material graph dataset, key to improving the generation quality. Finally, we generate high-quality graph samples using a regularized sampling process and improve the matching quality by differentiable optimization for top-ranked samples. We compare our methods to CLIP-based database search baselines (which are themselves novel) and achieve superior or similar performance without requiring massive data storage. We further show that our model can produce a set of material graphs unconditionally, conditioned on images, text prompts or partial graphs, serving as a tool for automatic visual programming completion.

We propose a surface-based cloth shading model that generates realistic cloth appearance with ply-level details. It generalizes previous surface-based models to a broader set of cloth including knitted and thin woven cloth. Our model takes into account the most dominant visual features of cloth, including anisotropic S-shaped reflection highlight, cross-shaped transmission highlights, delta transmission, and shadowing masking. We model these elements via a comprehensive micro-scale BSDF and a meso-scale effective BSDF formulation. Then, we propose an implementation that leverages the Monte Carlo sampler of path tracing for reducing precomputation to the bare minimum, by evaluating the effective BSDF as a Monte Carlo estimate, and encoding visibility using anisotropic spherical Gaussians. We demonstrate our model by replicating a set of woven and knitted fabrics, showing good match with respect to captured photographs.

We present a novel generative model, called Bidirectional GaitNet, that learns the relationship between human anatomy and its gait. The simulation model of human anatomy is a comprehensive, full-body, simulation-ready, musculoskeletal model with 304 Hill-type musculotendon units. The Bidirectional GaitNet consists of forward and backward models. The forward model predicts a gait pattern of a person with specific physical conditions, while the backward model estimates the physical conditions of a person when his/her gait pattern is provided. Our simulation-based approach first learns the forward model by distilling the simulation data generated by a state-of-the-art predictive gait simulator and then constructs a Variational Autoencoder (VAE) with the learned forward model as its decoder. Once it is learned its encoder serves as the backward model. We demonstrate our model on a variety of healthy/impaired gaits and validate it in comparison with physical examination data of real patients.

Great storytellers know how to take us on a journey. They direct characters to act—not necessarily in the most rational way—but rather in a way that leads to interesting situations, and ultimately creates an impactful experience for audience members looking on.

If audience experience is what matters most, then can we help artists and animators directly craft such experiences, independent of the concrete character actions needed to evoke those experiences? In this paper, we offer a novel computational framework for such tools. Our key idea is to optimize animations with respect to simulated audience members’ experiences. To simulate the audience, we borrow an established principle from cognitive science: that human social intuition can be modeled as “inverse planning,” the task of inferring an agent’s (hidden) goals from its (observed) actions. Building on this model, we treat storytelling as “inverse inverse planning,” the task of choosing actions to manipulate an inverse planner’s inferences. Our framework is grounded in literary theory, naturally capturing many storytelling elements from first principles. We give a series of examples to demonstrate this, with supporting evidence from human subject studies.

Creating pose-driven human avatars is about modeling the mapping from the low-frequency driving pose to high-frequency dynamic human appearances, so an effective pose encoding method that can encode high-fidelity human details is essential to human avatar modeling. To this end, we present PoseVocab, a novel pose encoding method that encourages the network to discover the optimal pose embeddings for learning the dynamic human appearance. Given multi-view RGB videos of a character, PoseVocab constructs key poses and latent embeddings based on the training poses. To achieve pose generalization and temporal consistency, we sample key rotations in so(3) of each joint rather than the global pose vectors, and assign a pose embedding to each sampled key rotation. These joint-structured pose embeddings not only encode the dynamic appearances under different key poses, but also factorize the global pose embedding into joint-structured ones to better learn the appearance variation related to the motion of each joint. To improve the representation ability of the pose embedding while maintaining memory efficiency, we introduce feature lines, a compact yet effective 3D representation, to model more fine-grained details of human appearances. Furthermore, given a query pose and a spatial position, a hierarchical query strategy is introduced to interpolate pose embeddings and acquire the conditional pose feature for dynamic human synthesis. Overall, PoseVocab effectively encodes the dynamic details of human appearance and enables realistic and generalized animation under novel poses. Experiments show that our method outperforms other state-of-the-art baselines both qualitatively and quantitatively in terms of synthesis quality. Code is available at https://github.com/lizhe00/PoseVocab.

This paper introduces DARAM, a dynamic avatar-human motion remapping technique that enables VR users to ascend virtual stairs. The primary design goal is to provide a realistic sensation of virtual stair walking while accounting for discrepancies between the user’s real body motion and the avatar’s motion, arising due to the virtual stairs present only in the virtual environment. Another design goal is to make DARAM applicable to dynamic multi-user environments. To this end, DARAM is designed to achieve motion remapping dynamically without requiring prior information about virtual stairs or environments, simplifying implementation in diverse VR applications. Furthermore, DARAM aims to synthesize avatar motion that delivers not only a realistic first-person experience but also a believable third-person experience for surrounding observers, making it applicable to multi-user VR applications. Two user studies demonstrate that the proposed technique successfully serves our design goals.

Daily objects embedded in a contextual environment are often ungraspable initially. Whether it is a book sandwiched by other books on a fully packed bookshelf or a piece of paper lying flat on the desk, a series of nonprehensile pregrasp maneuvers is required to manipulate the object into a graspable state. Humans are proficient at utilizing environmental contacts to achieve manipulation tasks that are otherwise impossible, but synthesizing such nonprehensile pregrasp behaviors is challenging to existing methods. We present a novel method that combines graph search, optimal control, and a learning-based objective function to synthesize physically realistic and diverse nonprehensile pre-grasp motions that leverage the external contacts. Since the “graspability” of an object in context with its surrounding is difficult to define, we utilize a dataset of dexterous grasps to learn a metric which implicitly takes into account the exposed surface of the object and the finger tip locations. Our method can efficiently discover hand and object trajectories that are certified to be physically feasible by the simulation and kinematically achievable by the dexterous hand. We evaluate our method on eight challenging scenarios where nonprehensile pre-grasps are required to succeed. We also show that our method can be applied to unseen objects different from those in the training dataset. Finally, we report quantitative analyses on generalization and robustness of our method, as well as an ablation study.

Large-scale text-to-image generative models have shown their remarkable ability to synthesize diverse, high-quality images. However, directly applying these models for real image editing remains challenging for two reasons. First, it is hard for users to craft a perfect text prompt depicting every visual detail in the input image. Second, while existing models can introduce desirable changes in certain regions, they often dramatically alter the input content and introduce unexpected changes in unwanted regions. In this work, we introduce pix2pix-zero, an image-to-image translation method that can preserve the original image’s content without manual prompting. We first automatically discover editing directions that reflect desired edits in the text embedding space. To preserve the content structure, we propose cross-attention guidance, which aims to retain the cross-attention maps of the input image throughout the diffusion process. Finally, to enable interactive editing, we distill the diffusion model into a fast conditional GAN. We conduct extensive experiments and show that our method outperforms existing and concurrent works for both real and synthetic image editing. In addition, our method does not need additional training for these edits and can directly use the existing pre-trained text-to-image diffusion model.

Text-to-image models (T2I) offer a new level of flexibility by allowing users to guide the creative process through natural language. However, personalizing these models to align with user-provided visual concepts remains a challenging problem. The task of T2I personalization poses multiple hard challenges, such as maintaining high visual fidelity while allowing creative control, combining multiple personalized concepts in a single image, and keeping a small model size. We present Perfusion, a T2I personalization method that addresses these challenges using dynamic rank-1 updates to the underlying T2I model. Perfusion avoids overfitting by introducing a new mechanism that “locks” new concepts’ cross-attention Keys to their superordinate category. Additionally, we develop a gated rank-1 approach that enables us to control the influence of a learned concept during inference time and to combine multiple concepts. This allows runtime efficient balancing of visual-fidelity and textual-alignment with a single 100KB trained model. Importantly, it can span different operating points across the Pareto front without additional training. We compare our approach to strong baselines and demonstrate its qualitative and quantitative strengths.

We introduce a novel regularization for localizing an elastic-energy-driven deformation to only those regions being manipulated by the user. Our local deformation features a natural region of influence, which is automatically adaptive to the geometry of the shape, the size of the deformation and the elastic energy in use. We further propose a three-block ADMM-based optimization to efficiently minimize the energy and achieve interactive frame rates. Our approach avoids the artifacts of other alternative methods, is simple and easy to implement, does not require tedious control primitive setup and generalizes across different dimensions and elastic energies. We demonstrates the effectiveness and efficiency of our localized deformation tool through a variety of local editing scenarios, including 1D, 2D, 3D elasticity and cloth deformation.

This paper focuses on how artificial intelligence (AI) can be used to assist general users in the creation of professional portraits, that is, consistently converting rough sketches into high-quality anime portraits during their sketching process. The input to this task is a sequence of incomplete human freehand sketches that are gradually refined stroke by stroke, while the output is a sequence of high-quality anime portraits that correspond to the input sketches as guidance. Although recent GANs can generate high quality images, it is a challenging problem to maintain the high quality of generated images from sketches with a low degree of completion due to ill-posed problems in conditional image generation. Even with the latest sketch-to-image (S2I) technology, it is still difficult to create high-quality images from incomplete rough sketches for anime portraits because the lines in anime style tend to be more abstract than in realistic style. In this paper, we addressed this problem using the latent space exploration of StyleGAN with a two-stage training strategy. Specifically, we consider the input strokes of a freehand sketch to correspond to edge information-related attributes in the latent structural code of StyleGAN, and term the matching between strokes and these attributes “stroke-level disentanglement.” In the first stage, we trained an image encoder with the pre-trained StyleGAN model as a teacher encoder. In the second stage, we simulated the drawing process of the generated images and trained the sketch encoder for incomplete progressive sketches to generate high-quality portrait images with feature alignment to the disentangled representations at the stroke level in the teacher encoder. We verified the proposed progressive S2I system with both qualitative and quantitative evaluations and achieved high-quality anime portraits from incomplete progressive sketches. What’s more, our user study proved its effectiveness in art creation assistance for the anime style.

Virtual reality (VR) passthrough uses external cameras on the front of a headset to allow the user to see their environment. However, passthrough cameras cannot physically be co-located with the user’s eyes, so the passthrough images have a different perspective than what the user would see without the headset. Although the images can be computationally reprojected into the desired view, errors in depth estimation, view-dependent effects, and missing information at occlusion boundaries can lead to undesirable artifacts.

We propose a novel computational camera that directly samples the rays that would have gone into the user’s eye, several centimeters behind the sensor. Our design contains an array of lenses with an aperture behind each lens, and the apertures are strategically placed to allow through only the desired rays. The resulting thin, flat architecture has suitable form factor for VR, and the image reconstruction is computationally lightweight, enabling low-latency passthrough. We demonstrate our approach experimentally in a fully functional binocular passthrough prototype with practical calibration and real-time image reconstruction. Finally, we experimentally validate that our camera captures the correct perspective for VR passthrough, even in the presence of transparent objects, specular highlights, and complex occluding structures.

We present the UrbanBIS benchmark for large-scale 3D urban understanding, supporting practical urban-level semantic and building-level instance segmentation. UrbanBIS comprises six real urban scenes, with 2.5 billion points, covering a vast area of 10.78 km2 and 3,370 buildings, captured by 113,346 views of aerial photogrammetry. Particularly, UrbanBIS provides not only semantic-level annotations on a rich set of urban objects, including buildings, vehicles, vegetation, roads, and bridges, but also instance-level annotations on the buildings. Further, UrbanBIS is the first 3D dataset that introduces fine-grained building sub-categories, considering a wide variety of shapes for different building types. Besides, we propose B-Seg, a building instance segmentation method to establish UrbanBIS. B-Seg adopts an end-to-end framework with a simple yet effective strategy for handling large-scale point clouds. Compared with mainstream methods, B-Seg achieves better accuracy with faster inference speed on UrbanBIS. In addition to the carefully-annotated point clouds, UrbanBIS provides high-resolution aerial-acquisition photos and high-quality large-scale 3D reconstruction models, which shall facilitate a wide range of studies such as multi-view stereo, urban LOD generation, aerial path planning, autonomous navigation, road network extraction, and so on, thus serving as an important platform for many intelligent city applications. UrbanBIS and related code can be downloaded at https://vcc.tech/UrbanBIS.

Inverse rendering methods that account for global illumination are becoming more popular, but current methods require evaluating and automatically differentiating millions of path integrals by tracing multiple light bounces, which remains expensive and prone to noise. Instead, this paper proposes a radiometric prior as a simple alternative to building complete path integrals in a traditional differentiable path tracer, while still correctly accounting for global illumination. Inspired by the Neural Radiosity technique, we use a neural network as a radiance function, and we introduce a prior consisting of the norm of the residual of the rendering equation in the inverse rendering loss. We train our radiance network and optimize scene parameters simultaneously using a loss consisting of both a photometric term between renderings and the multi-view input images, and our radiometric prior (the residual term). This residual term enforces a physical constraint on the optimization that ensures that the radiance field accounts for global illumination. We compare our method to a vanilla differentiable path tracer, and more advanced techniques such as Path Replay Backpropagation. Despite the simplicity of our approach, we can recover scene parameters with comparable and in some cases better quality, at considerably lower computation times.

Differentiable rendering is frequently used in gradient descent-based inverse rendering pipelines to solve for scene parameters – such as reflectance or lighting properties – from target image inputs. Efficient computation of accurate, low variance gradients is critical for rapid convergence. While many methods employ variance reduction strategies, they operate independently on each gradient descent iteration, requiring large sample counts and computation. Gradients may however vary slowly between iterations, leading to unexplored potential benefits when reusing sample information to exploit this coherence. We develop an algorithm to reuse Monte Carlo gradient samples between gradient iterations, motivated by reservoir-based temporal importance resampling in forward rendering. Direct application of this method is not feasible, as we are computing many derivative estimates (i.e., one per optimization parameter) instead of a single pixel intensity estimate; moreover, each of these gradient estimates can affect multiple pixels, and gradients can take on negative values. We address these challenges by reformulating differential rendering integrals in parameter space, developing a new resampling estimator that treats negative functions, and combining these ideas into a reuse algorithm for inverse texture optimization. We significantly reduce gradient error compared to baselines, and demonstrate faster inverse rendering convergence in settings involving complex direct lighting and material textures.

We investigate the problem of accelerating a physically-based differentiable renderer for heightfields based on path tracing with global illumination. On a heightfield with 1 million vertices (1024 × 1024 resolution), our differentiable renderer requires only 4 ms per sample per pixel when differentiating direct illumination, orders of magnitude faster than most existing general 3D mesh differentiable renderers. It is well-known that one can leverage spatial hierarchical data structures (e.g., the maximum mipmaps) to accelerate the forward pass of heightfield rendering. The key idea of our approach is to further utilize the hierarchy to speed up the backward pass—differentiable heightfield rendering. Specifically, we use the maximum mipmaps to accelerate the process of identifying scene discontinuities, which is crucial for obtaining accurate derivatives. Our renderer supports global illumination. we are able to optimize global effects, such as shadows, with respect to the geometry and the material parameters. Our differentiable renderer achieves real-time frame rates and unlocks interactive inverse rendering applications. We demonstrate the flexibility of our method with terrain optimization, geometric illusions, shadow optimization, and text-based shape generation.

We present a technique to optimize the reflectivity of a surface while preserving its overall shape. The naïve optimization of the mesh vertices using the gradients of reflectivity simulations results in undesirable distortion. In contrast, our robust formulation optimizes the surface normal as an independent variable that bridges the reflectivity term with differential rendering, and the regularization term with as-rigid-as-possible elastic energy. We further adaptively subdivide the input mesh to improve the convergence. Consequently, our method can minimize the retroreflectivity of a wide range of input shapes, resulting in sharply creased shapes ubiquitous among stealth aircraft and Sci-Fi vehicles. Furthermore, by changing the reward for the direction of the outgoing light directions, our method can be applied to other reflectivity design tasks, such as the optimization of architectural walls to concentrate light in a specific region. We have tested the proposed method using light-transport simulations and real-world 3D-printed objects.

Color and gloss are fundamental aspects of surface appearance. State-of-the-art fabrication techniques can manipulate both properties of the printed 3D objects. However, in the context of appearance reproduction, perceptual aspects of color and gloss are usually handled separately, even though previous perceptual studies suggest their interaction. Our work is motivated by previous studies demonstrating a perceived color shift due to a change in the object’s gloss, i.e., two samples with the same color but different surface gloss appear as they have different colors. In this paper, we conduct new experiments which support this observation and provide insights into the magnitude and direction of the perceived color change. We use the observations as guidance to design a new method that estimates and corrects the color shift enabling the fabrication of objects with the same perceived color but different surface gloss. We formulate the problem as an optimization procedure solved using differentiable rendering. We evaluate the effectiveness of our method in perceptual experiments with 3D objects fabricated using a multi-material 3D printer and demonstrate potential applications.

Many computational algorithms applied to geometry operate on discrete representations of shape. It is sometimes necessary to first simplify, or coarsen, representations found in modern datasets for practicable or expedited processing. The utility of a coarsening algorithm depends on both, the choice of representation as well as the specific processing algorithm or operator. e.g. simulation using the Finite Element Method, calculating Betti numbers, etc. We propose a novel method that can coarsen triangle meshes, tetrahedral meshes and simplicial complexes. Our method allows controllable preservation of salient features from the high-resolution geometry and can therefore be customized to different applications.

Salient properties are typically captured by local shape descriptors via linear differential operators – variants of Laplacians. Eigenvectors of their discretized matrices yield a useful spectral domain for geometry processing (akin to the famous Fourier spectrum which uses eigenfunctions of the derivative operator). Existing methods for spectrum-preserving coarsening use zero-dimensional discretizations of Laplacian operators (defined on vertices). We propose a generalized spectral coarsening method that considers multiple Laplacian operators defined in different dimensionalities in tandem. Our simple algorithm greedily decides the order of contractions of simplices based on a quality function per simplex. The quality function quantifies the error due to removal of that simplex on a chosen band within the spectrum of the coarsened geometry.

Cage coordinates are a powerful means to define 2D deformation fields from sparse control points. We introduce Conformal polynomial Coordinates for closed polyhedral cages, enabling segments to be transformed into polynomial curves of any order. Extending classical 2D Green coordinates, our coordinates result in conformal harmonic deformations that are cage-aware. We demonstrate the usefulness of our technique on a variety of 2D deformation scenarios where curves allow artists to perform intuitive deformations with few input parameters. Our method combines the texture preservation property of conformal deformations together with the expressiveness offered by Bezier controls.

We devise a local–global solver dedicated to the simulation of Discrete Elastic Rods (DER) with Coulomb friction that can fully leverage the massively parallel compute capabilities of moderns GPUs. We verify that our simulator can reproduce analytical results on recently published cantilever, bend–twist, and stick–slip experiments, while drastically decreasing iteration times for high-resolution hair simulations. Being able to handle contacting assemblies of several thousand elastic rods in real-time, our fast solver paves the ways for new workflows such as interactive physics-based editing of digital grooms.

We present a novel mesh-based method for simulating the intricate dynamics of (potentially multi-layered) continuum thick shells. In order to accurately represent the constitutive behavior of structural responses in the thickness direction, we develop a dual-quadrature prism finite element formulation that is free from shear locking and naturally incorporates three-dimensional elastoplastic and viscoelastic constitutive models. Additionally, we introduce a simple and effective technique for coupling a high-resolution membrane layer on top of the thick shell to enable complementary high-frequency deformation modes that generate realistic wrinkles. With our novelly designed sparse basis vectors for the high-frequency deformations, the constrained Lagrangian mechanics problem is expressed as an unconstrained optimization and then efficiently solved by a custom alternating minimization technique. Our method opens up a new possibility for fast, high-quality, and thickness-aware simulations of leather garments, pillows, mats, metal boards, and potentially a variety of other thick structures.

Stable, low-cost, and precise visual measurement of directional information has many applications in domains such as virtual and augmented reality, visual odometry, or industrial computer vision. Conventional approaches like checkerboard patterns require careful pre-calibration, and can therefore not be operated in snapshot mode. Other optical methods like autocollimators offer very high precision but require controlled environments and are hard to take outside the lab. Non-optical methods like IMUs are low cost and widely available, but suffer from high drift errors.

To overcome these challenges, we propose a novel snapshot method for angular measurement and tracking with Moiré patterns that are generated by binary structures printed on both sides of a glass plate. The Moiré effect amplifies minute angular shifts and translates them into spatial phase shifts that can be readily measured with a camera, effectively implementing an optical Vernier scale. We further extend this principle from a simple phase shift to a chirp model, which allows for full 6D tracking as well as estimation of camera intrinsics like the field of view. Simulation and experimental results show that the proposed non-contact object tracking framework is computationally efficient and the average angular accuracy of 0.17° outperforms the state-of-the-arts.

Ergonomic efficiency is essential to the mass and prolonged adoption of VR/AR experiences. While VR/AR head-mounted displays unlock users’ natural wide-range head movements during viewing, their neck muscle comfort is inevitably compromised by the added hardware weight. Unfortunately, little quantitative knowledge for understanding and addressing such an issue is available so far.

Leveraging electromyography devices, we measure, model, and predict VR users’ neck muscle contraction levels (MCL) while they move their heads to interact with the virtual environment. Specifically, by learning from collected physiological data, we develop a bio-physically inspired computational model to predict neck MCL under diverse head kinematic states. Beyond quantifying the cumulative MCL of completed head movements, our model can also predict potential MCL requirements with target head poses only. A series of objective evaluations and user studies demonstrate its prediction accuracy and generality, as well as its ability in reducing users’ neck discomfort by optimizing the layout of visual targets. We hope this research will motivate new ergonomic-centered designs for VR/AR and interactive graphics applications. Source code is released at: https://github.com/NYU-ICL/xr-ergonomics-neck-comfort.

We introduce a wearable single-eye emotion recognition device and a real-time approach to recognizing emotions from partial observations of an emotion that is robust to changes in lighting conditions. At the heart of our method is a bio-inspired event-based camera setup and a newly designed lightweight Spiking Eye Emotion Network (SEEN). Compared to conventional cameras, event-based cameras offer a higher dynamic range (up to 140 dB vs. 80 dB) and a higher temporal resolution (in the order of μ s vs. 10s of ms). Thus, the captured events can encode rich temporal cues under challenging lighting conditions. However, these events lack texture information, posing problems in decoding temporal information effectively. SEEN tackles this issue from two different perspectives. First, we adopt convolutional spiking layers to take advantage of the spiking neural network’s ability to decode pertinent temporal information. Second, SEEN learns to extract essential spatial cues from corresponding intensity frames and leverages a novel weight-copy scheme to convey spatial attention to the convolutional spiking layers during training and inference. We extensively validate and demonstrate the effectiveness of our approach on a specially collected Single-eye Event-based Emotion (SEE) dataset. To the best of our knowledge, our method is the first eye-based emotion recognition method that leverages event-based cameras and spiking neural networks.

Previous path guiding techniques typically rely on spatial subdivision structures to approximate directional target distributions, which may cause failure to capture spatio-directional correlations and introduce parallax issue. In this paper, we present Neural Parametric Mixtures (NPM), a neural formulation to encode target distributions for path guiding algorithms. We propose to use a continuous and compact neural implicit representation for encoding parametric models while decoding them via lightweight neural networks. We then derive a gradient-based optimization strategy to directly train the parameters of NPM with noisy Monte Carlo radiance estimates. Our approach efficiently models the target distribution (incident radiance or the product integrand) for path guiding, and outperforms previous guiding methods by capturing the spatio-directional correlations more accurately. Moreover, our approach is more training efficient and is practical for parallelization on modern GPUs.

Focal points are fascinating effects that emerge from various constellations, for example when light passes through narrow gaps or when objects are seen through lenses or mirrors. These effects can be challenging to render, as paths need to pass through small regions that are not always known beforehand and can occur freely in space. Specialized algorithms exist for some effects, but many of them rely on Markov chain Monte Carlo integration, which is known to suffer from uneven convergence undesirable in practice. Path guiding methods are a promising alternative, but existing techniques only handle a subset of focal effects. We propose a novel form of guiding that is specifically tailored to identify focal points and sample them in accordance to their image contribution. Our technique is the first to unify all focal effects in a single framework and we demonstrate that it can render effects that previous state-of-the-art techniques are unable to handle.

Null-collision approaches for estimating transmittance and sampling free-flight distances are the current state-of-the-art for unbiased rendering of general heterogeneous participating media. However, null-collision approaches have a strict requirement for specifying a tightly bounding total extinction in order to remain both robust and performant; in practice this requirement restricts the use of null-collision techniques to only participating media where the density of the medium at every possible point in space is known a-priori. In production rendering, a common case is a medium in which density is defined by a black-box procedural function for which a bounding extinction cannot be determined beforehand. Typically in this case, a bounding extinction must be approximated by using an overly loose and therefore computationally inefficient conservative estimate. We present an analysis of how null-collision techniques degrade when a more aggressive initial guess for a bounding extinction underestimates the true maximum density and turns out to be non-bounding. We then build upon this analysis to arrive at two new techniques: first, a practical, efficient, consistent progressive algorithm that allows us to robustly adapt null-collision techniques for use with procedural media with unknown bounding extinctions, and second, a new importance sampling technique that improves ratio-tracking based on zero-variance sampling.

Monte Carlo rendering is a computationally intensive task, but combined with recent deep-learning based advances in image denoising it is possible to achieve high quality images in a shorter amount of time. We present a novel adaptive sampling technique that further improves the efficiency of Monte Carlo rendering combined with deep-learning based denoising. Our proposed technique is general, can be combined with existing pre-trained denoisers, and, in contrast with previous techniques, does not itself require any additional neural networks or learning. A key contribution of our work is a general method for estimating the variance of the outputs of a neural network whose inputs are random variables. Our method iteratively renders additional samples and uses this novel variance estimate to compute the sample distribution for each subsequent iteration. Compared to uniform sampling and previous adaptive sampling techniques, our method achieves better equal-time error in all scenes tested, and when combined with a recent denoising post-correction technique, significantly faster error convergence is realized.

In this paper, we focus on the task of 3D shape completion from partial point clouds using deep implicit functions. Existing methods seek to use voxelized basis functions or the ones from a certain family of functions (e.g., Gaussians), which leads to high computational costs or limited shape expressivity. On the contrary, our method employs adaptive local basis functions, which are learned end-to-end and not restricted in certain forms. Based on those basis functions, a local-to-local shape completion framework is presented. Our algorithm learns sparse parameterization with a small number of basis functions while preserving local geometric details during completion. Quantitative and qualitative experiments demonstrate that our method outperforms the state-of-the-art methods in shape completion, detail preservation, generalization to unseen geometries, and computational cost. Code and data for this paper are at https://github.com/yinghdb/Adaptive-Local-Basis-Functions.

We derive a minimalist but powerful deterministic denoising-diffusion model. While denoising diffusion has shown great success in many domains, its underlying theory remains largely inaccessible to non-expert users. Indeed, an understanding of graduate-level concepts such as Langevin dynamics or score matching appears to be required to grasp how it works. We propose an alternative approach that requires no more than undergrad calculus and probability. We consider two densities and observe what happens when random samples from these densities are blended (linearly interpolated). We show that iteratively blending and deblending samples produces random paths between the two densities that converge toward a deterministic mapping. This mapping can be evaluated with a neural network trained to deblend samples. We obtain a model that behaves like deterministic denoising diffusion: it iteratively maps samples from one density (e.g., Gaussian noise) to another (e.g., cat images). However, compared to the state-of-the-art alternative, our model is simpler to derive, simpler to implement, more numerically stable, achieves higher quality results in our experiments, and has interesting connections to computer graphics.

We present multimodal conditioning modules (MCM) for enabling conditional image synthesis using pretrained diffusion models. Previous multimodal synthesis works rely on training networks from scratch or fine-tuning pretrained networks, both of which are computationally expensive for large, state-of-the-art diffusion models. Our method uses pretrained networks but does not require any updates to the diffusion network’s parameters. MCM is a small module trained to modulate the diffusion network’s predictions during sampling using 2D modalities (e.g., semantic segmentation maps, sketches) that were unseen during the original training of the diffusion model. We show that MCM enables user control over the spatial layout of the image and leads to increased control over the image generation process. Training MCM is cheap as it does not require gradients from the original diffusion net, consists of only ∼ 1% of the number of parameters of the base diffusion model, and is trained using only a limited number of training examples. We evaluate our method on unconditional and text-conditional models to demonstrate the improved control over the generated images and their alignment with respect to the conditioning inputs.

Realistic, scalable, and controllable generation of furniture layouts is essential for many applications in virtual reality, augmented reality, game development and synthetic data generation. The most successful current methods tackle this problem as a sequence generation problem which imposes a specific ordering on the elements of the layout, making it hard to exert fine-grained control over the attributes of a generated scene. Existing methods provide control through object-level conditioning, or scene completion, where generation can be conditioned on an arbitrary subset of furniture objects. However, attribute-level conditioning, where generation can be conditioned on an arbitrary subset of object attributes, is not supported. We propose COFS, a method to generate furniture layouts that enables fine-grained control through attribute-level conditioning. For example, COFS allows specifying only the scale and type of objects that should be placed in the scene and the generator chooses their positions and orientations; or the position that should be occupied by objects can be specified and the generator chooses their type, scale, orientation, etc. Our results show both qualitatively and quantitatively that we significantly outperform existing methods on attribute-level conditioning.

In this work, we present Conditional Adversarial Latent Models  (CALM), an approach for generating diverse and directable behaviors for user-controlled interactive virtual characters. Using imitation learning, CALM  learns a representation of movement that captures the complexity and diversity of human motion, and enables direct control over character movements. The approach jointly learns a control policy and a motion encoder that reconstructs key characteristics of a given motion without merely replicating it. The results show that CALM  learns a semantic motion representation, enabling control over the generated motions and style-conditioning for higher-level task training. Once trained, the character can be controlled using intuitive interfaces, akin to those found in video games.

Styled online in-between motion generation has important application scenarios in computer animation and games. Its core challenge lies in the need to satisfy four critical requirements simultaneously: generation speed, motion quality, style diversity, and synthesis controllability. While the first two challenges demand a delicate balance between simple fast models and learning capacity for generation quality, the latter two are rarely investigated together in existing methods, which largely focus on either control without style or uncontrolled stylized motions. To this end, we propose a Real-time Stylized Motion Transition method (RSMT) to achieve all aforementioned goals. Our method consists of two critical, independent components: a general motion manifold model and a style motion sampler. The former acts as a high-quality motion source and the latter synthesizes styled motions on the fly under control signals. Since both components can be trained separately on different datasets, our method provides great flexibility, requires less data, and generalizes well when no/few samples are available for unseen styles. Through exhaustive evaluation, our method proves to be fast, high-quality, versatile, and controllable. The code and data are available at https://github.com/yuyujunjun/RSMT-Realtime-Stylized-Motion-Transition.

Computing occluding contours is a key step in 3D non-photorealistic rendering, but producing smooth contours with consistent visibility has been a notoriously-challenging open problem. This paper describes the first general-purpose smooth surface construction for which the occluding contours can be computed in closed form. Given an input mesh and camera viewpoint, we show how to approximate the mesh with a G1 piecewise-quadratic surface, for which the occluding contours are piecewise-rational curves in image-space. We show that this method produces smooth contours with consistent visibility much more efficiently than the state-of-the-art.

Physical systems ranging from elastic bodies to kinematic linkages are defined on high-dimensional configuration spaces, yet their typical low-energy configurations are concentrated on much lower-dimensional subspaces. This work addresses the challenge of identifying such subspaces automatically: given as input an energy function for a high-dimensional system, we produce a low-dimensional map whose image parameterizes a diverse yet low-energy submanifold of configurations. The only additional input needed is a single seed configuration for the system to initialize our procedure; no dataset of trajectories is required. We represent subspaces as neural networks that map a low-dimensional latent vector to the full configuration space, and propose a training scheme to fit network parameters to any system of interest. This formulation is effective across a very general range of physical systems; our experiments demonstrate not only nonlinear and very low-dimensional elastic body and cloth subspaces, but also more general systems like colliding rigid bodies and linkages. We briefly explore applications built on this formulation, including manipulation, latent interpolation, and sampling.

Neural material representations have recently been proposed to augment the material appearance toolbox used in realistic rendering. These models are successful at tasks ranging from measured BTF compression, through efficient rendering of synthetic displaced materials with occlusions, to BSDF layering. However, importance sampling has been an after-thought in most neural material approaches, and has been handled by inefficient cosine-hemisphere sampling or mixing it with an additional simple analytic lobe. In this paper we fill that gap, by evaluating and comparing various pdf-learning approaches for sampling spatially varying neural materials, and proposing new variations of these approaches. We investigate three sampling approaches: analytic-lobe mixtures, normalizing flows, and histogram prediction. Within each type, we introduce improvements beyond previous work, and we extensively evaluate and compare these approaches in terms of sampling rate, wall-clock time, and final visual quality. Our versions of normalizing flows and histogram mixtures perform well and can be used in practical rendering systems, potentially facilitating the broader adoption of neural material models in production.

Spectral rendering is a crucial solution for photorealistic rendering. However, most available texture assets are RGB-only, and access to spectral content is limited. Uplifting methods that recover full spectral representations from RGB inputs have therefore received much attention. Yet, most methods are deterministic, while, in reality, there is no one-to-one mapping. As a consequence, the appearance of uplifted textures is fully determined under all illuminants. Hereby, metamers, which are materials with differing spectral responses that appear identical under a specific illumination, are excluded.

We propose a method which makes this uplifting process controllable. Hereby, a user can define texture appearance under various lighting conditions, leading to a greatly increased flexibility for content design. Our method determines the space of possible metameric manipulations and enables interactive adjustments, while maintaining a set of user-specified appearance constraints. To achieve this goal, we formulate the problem as a constrained optimization, building upon an interpolation scheme and data-based reflectance generation, which maintain plausibility. Besides its value for artistic control, our solution is lightweight and can be executed on the fly, which keeps its memory consumption low and makes it easy to integrate into existing frameworks.

Texture synthesis is a fundamental problem in computer graphics that would benefit various applications. Existing methods are effective in handling 2D image textures. In contrast, many real-world textures contain meso-structure in the 3D geometry space, such as grass, leaves, and fabrics, which cannot be effectively modeled using only 2D image textures. We propose a novel texture synthesis method with Neural Radiance Fields (NeRF) to capture and synthesize textures from given multi-view images. In the proposed NeRF texture representation, a scene with fine geometric details is disentangled into the meso-structure textures and the underlying base shape. This allows textures with meso-structure to be effectively learned as latent features situated on the base shape, which are fed into a NeRF decoder trained simultaneously to represent the rich view-dependent appearance. Using this implicit representation, we can synthesize NeRF-based textures through patch matching of latent features. However, inconsistencies between the metrics of the reconstructed content space and the latent feature space may compromise the synthesis quality. To enhance matching performance, we further regularize the distribution of latent features by incorporating a clustering constraint. Experimental results and evaluations demonstrate the effectiveness of our approach.

Fitting primitives for point cloud data to obtain a structural representation has been widely adopted for reverse engineering and other graphics applications. Existing segmentation-based approaches only segment primitive patches but ignore edges that indicate boundaries of primitives, leading to inaccurate and incomplete reconstruction. To fill the gap, we present a novel surface and edge detection network (SED-Net) for accurate geometric primitive fitting of point clouds. The key idea is to learn parametric surfaces (including B-spline patches) and edges jointly that can be assembled into a regularized and seamless CAD model in one unified and efficient framework. SED-Net is equipped with a two-branch structure to extract type and edge features and geometry features of primitives. At the core of our network is a two-stage feature fusion mechanism to utilize the type, edge and geometry features fully. Precisely detected surface patches can be employed as contextual information to facilitate the detection of edges and corners. Benefiting from the simultaneous detection of surfaces and edges, we can obtain a parametric and compact model representation. This enables us to represent a CAD model with predefined primitive-specific meshes and also allows users to edit its shape easily. Extensive experiments and comparisons against previous methods demonstrate our effectiveness and superiority.

Inspired by the strengths of quadric error metrics initially designed for mesh decimation, we propose a concise mesh reconstruction approach for 3D point clouds. Our approach proceeds by clustering the input points enriched with quadric error metrics, where the generator of each cluster is the optimal 3D point for the sum of its quadric error metrics. This approach favors the placement of generators on sharp features, and tends to equidistribute the error among clusters. We reconstruct the output surface mesh from the adjacency between clusters and a constrained binary solver. We combine our clustering process with an adaptive refinement driven by the error. Compared to prior art, our method avoids dense reconstruction prior to simplification and produces immediately an optimized mesh.

We present a method for reconstructing high-quality meshes of large unbounded real-world scenes suitable for photorealistic novel view synthesis. We first optimize a hybrid neural volume-surface scene representation designed to have well-behaved level sets that correspond to surfaces in the scene. We then bake this representation into a high-quality triangle mesh, which we equip with a simple and fast view-dependent appearance model based on spherical Gaussians. Finally, we optimize this baked representation to best reproduce the captured viewpoints, resulting in a model that can leverage accelerated polygon rasterization pipelines for real-time view synthesis on commodity hardware. Our approach outperforms previous scene representations for real-time rendering in terms of accuracy, speed, and power consumption, and produces high quality meshes that enable applications such as appearance editing and physical simulation.

With NeRF widely used for facial reenactment, recent methods can recover photo-realistic 3D head avatar from just a monocular video. Unfortunately, the training process of the NeRF-based methods is quite time-consuming, as MLP used in the NeRF-based methods is inefficient and requires too many iterations to converge. To overcome this problem, we propose AvatarMAV, a fast 3D head avatar reconstruction method using Motion-Aware Neural Voxels. AvatarMAV is the first to model both the canonical appearance and the decoupled expression motion by neural voxels for head avatar. In particular, the motion-aware neural voxels is generated from the weighted concatenation of multiple 4D tensors. The 4D tensors semantically correspond one-to-one with 3DMM expression basis and share the same weights as 3DMM expression coefficients. Benefiting from our novel representation, the proposed AvatarMAV can recover photo-realistic head avatars in just 5 minutes (implemented with pure PyTorch), which is significantly faster than the state-of-the-art facial reenactment methods. Project page: https://www.liuyebin.com/avatarmav.

Estimating spatially varying BRDF from a single image without complicated acquisition devices is a challenging problem. In this paper, a deep learning based method was proposed to improve the capture efficiency of single image significantly by learning the lighting pattern of a planar light source, and reconstruct high-quality SVBRDF by learning the global correlation prior of the input image. In our framework, the lighting pattern optimization is embedded in the training process of the network by introducing an online rendering process. The rendering process not only renders images online as the input of network, but also efficiently back propagates gradients from the network to optimize the lighting pattern. Once trained, the network can estimate SVBRDFs from real photographs captured under the learned lighting pattern. Additionally, we describe an onsite capture setup that needs no careful calibration to capture the material sample efficiently. In particular, even a cell phone can be used for illumination. We demonstrate on synthetic and real data that our method could recover a wide range of materials from a single image casually captured under the learned lighting pattern.

Authoring high-quality digital materials is key to realism in 3D rendering. Previous generative models for materials have been trained exclusively on synthetic data; such data is limited in availability and has a visual gap to real materials. We circumvent this limitation by proposing PhotoMat: the first material generator trained exclusively on real photos of material samples captured using a cell phone camera with flash. Supervision on individual material maps is not available in this setting. Instead, we train a generator for a neural material representation that is rendered with a learned relighting module to create arbitrarily lit RGB images; these are compared against real photos using a discriminator. We train PhotoMat with a new dataset of 12,000 material photos captured with handheld phone cameras under flash lighting. We demonstrate that our generated materials have better visual quality than previous material generators trained on synthetic data. Moreover, we can fit analytical material models to closely match these generated neural materials, thus allowing for further editing and use in 3D rendering.

Three-dimensional mesostructures enrich coarse macrosurfaces with complex features, which are 3D geometry with arbitrary topology in essence, but are expected to be self-similar with no tiling artifacts, just like texture-based material models. This is a challenging task, as no existing modeling tool provides the right constraints in the design phase to ensure such properties while maintaining real-time editing capabilities. In this paper, we propose MesoGen, a novel tile-centric authoring approach for the design of procedural mesostructures featuring non-periodic self-similarity while being represented as a compact and GPU-friendly model. We ensure by construction the continuity of the mesostructure: the user designs a set of atomic tiles by drawing 2D cross-sections on the interfaces between tiles, and selecting pairs of cross-sections to be connected as strands, i.e., 3D sweep surfaces. In parallel, a tiling engine continuously fills the shell space of the macrosurface with the so-defined tile set while ensuring that only matching interfaces are in contact. Moreover, the engine suggests to the user the addition of new tiles whenever the problem happens to be over-constrained. As a result, our method allows for the rapid creation of complex, seamless procedural mesostructure and is particularly adapted for wicker-like ones, often impossible to achieve with scattering-based mesostructure synthesis methods.

We present a method for the digital fabrication of surfaces whose appearance varies based on viewing direction. The surfaces are constructed from a mesh of bars arranged in a self-occluding colored heightfield that creates the desired view-dependent effects. At the heart of our method is a novel and simple differentiable rendering algorithm specifically designed to render colored 3D heightfields and enable efficient calculation of the gradient of appearance with respect to heights and colors. This algorithm forms the basis of a coarse-to-fine ML-based optimization process that adjusts the heights and colors of the strips to minimize the loss between the desired and real surface appearance from each viewpoint, deriving meshes that can then be fabricated using a 3D printer. Using our method, we demonstrate both synthetic and real-world fabricated results with view-dependent appearance.

In this paper, we present a novel cage deformer based on elasticity-derived matrix-valued coordinates. In order to bypass the typical shearing artifacts and lack of volume control of existing cage deformers, we promote a more elastic behavior of the cage deformation by deriving our coordinates from the Somigliana identity, a boundary integral formulation based on the fundamental solution of linear elasticity. Given an initial cage and its deformed pose, the deformation of the cage interior is deduced from these Somigliana coordinates via a corotational scheme, resulting in a matrix-weighted combination of both vertex positions and face normals of the cage. Our deformer thus generalizes Green coordinates, while producing physically-plausible spatial deformations that are invariant under similarity transformations and with interactive bulging control. We demonstrate the efficiency and versatility of our method through a series of examples in 2D and 3D.

In 3D shape reconstruction based on template mesh deformation, a regularization, such as smoothness energy, is employed to guide the reconstruction into a desirable direction. In this paper, we highlight an often overlooked property in the regularization: the vertex density in the mesh. Without careful control on the density, the reconstruction may suffer from under-sampling of vertices near shape details. We propose a novel mesh density adaptation method to resolve the under-sampling problem. Our mesh density adaptation energy increases the density of vertices near complex structures via deformation to help reconstruction of shape details. We demonstrate the usability and performance of mesh density adaptation with two tasks, inverse rendering and non-rigid surface registration. Our method produces more accurate reconstruction results compared to the cases without mesh density adaptation. Our code is available at https://github.com/ycjungSubhuman/density-adaptation.

In this paper, we present TEXTure, a novel method for text-guided generation, editing, and transfer of textures for 3D shapes. Leveraging a pretrained depth-to-image diffusion model, TEXTure applies an iterative scheme that paints a 3D model from different viewpoints. Yet, while depth-to-image models can create plausible textures from a single viewpoint, the stochastic nature of the generation process can cause many inconsistencies when texturing an entire 3D object. To tackle these problems, we dynamically define a trimap partitioning of the rendered image into three progression states, and present a novel elaborated diffusion sampling process that uses this trimap representation to generate seamless textures from different views. We then show that one can transfer the generated texture maps to new 3D geometries without requiring explicit surface-to-surface mapping, as well as extract semantic textures from a set of images without requiring any explicit reconstruction. Finally, we show that TEXTure can be used to not only generate new textures but also edit and refine existing textures using either a text prompt or user-provided scribbles. We demonstrate that our TEXTuring method excels at generating, transferring, and editing textures through extensive evaluation, and further close the gap between 2D image generation and 3D texturing. Code is available via our project page: https://texturepaper.github.io/TEXTurePaper/.

Text-to-Image models have introduced a remarkable leap in the evolution of machine learning, demonstrating high-quality synthesis of images from a given text-prompt. However, these powerful pretrained models still lack control handles that can guide spatial properties of the synthesized images. In this work, we introduce a universal approach to guide a pretrained text-to-image diffusion model, with a spatial map from another domain (e.g., sketch) during inference time. Unlike previous works, our method does not require to train a dedicated model or a specialized encoder for the task. Our key idea is to train a Latent Guidance Predictor (LGP) - a small, per-pixel, Multi-Layer Perceptron (MLP) that maps latent features of noisy images to spatial maps, where the deep features are extracted from the core Denoising Diffusion Probabilistic Model (DDPM) network. The LGP is trained only on a few thousand images and constitutes a differential guiding map predictor, over which the loss is computed and propagated back to push the intermediate images to agree with the spatial map. The per-pixel training offers flexibility and locality which allows the technique to perform well on out-of-domain sketches, including free-hand style drawings. We take a particular focus on the sketch-to-image translation task, revealing a robust and expressive way to generate images that follow the guidance of a sketch of arbitrary style or domain.

Virtual try-on attracts increasing research attention as a promising way for enhancing the user experience for online cloth shopping. Though existing methods can generate impressive results, users need to provide a well-designed reference image containing the target fashion clothes that often do not exist. To support user-friendly fashion customization in full-body portraits, we propose a multi-modal interactive setting by combining the advantages of both text and texture for multi-level fashion manipulation. With the carefully designed fashion editing module and loss functions, FashionTex framework can semantically control cloth types and local texture patterns without annotated pairwise training data. We further introduce an ID recovery module to maintain the identity of input portrait. Extensive experiments have demonstrated the effectiveness of our proposed pipeline. Code for this paper are at https://github.com/picksh/FashionTex.

Recently introduced Contrastive Language-Image Pre-Training (CLIP) [Radford et al. 2021] bridges images and text by embedding them into a joint latent space. This opens the door to ample literature that aims to manipulate an input image by providing a textual explanation. However, due to the discrepancy between image and text embeddings in the joint space, using text embeddings as the optimization target often introduces undesired artifacts in the resulting images. Disentanglement, interpretability, and controllability are also hard to guarantee for manipulation. To alleviate these problems, we propose to define corpus subspaces spanned by relevant prompts to capture specific image characteristics. We introduce CLIP projection-augmentation embedding (PAE) as an optimization target to improve the performance of text-guided image manipulation. Our method is a simple and general paradigm that can be easily computed and adapted, and smoothly incorporated into any CLIP-based image manipulation algorithm. To demonstrate the effectiveness of our method, we conduct several theoretical and empirical studies. As a case study, we utilize the method for text-guided semantic face editing. We quantitatively and qualitatively demonstrate that PAE facilitates a more disentangled, interpretable, and controllable face image manipulation with state-of-the-art quality and accuracy.

Finding correspondences between shapes is a central task in geometry processing with applications such as texture or deformation transfer and shape interpolation. We develop a spectral method for finding correspondences between non-isometric shapes that aligns extrinsic features. For this, we propose a novel crease aware spectral basis, that is derived from the Hessian of an elastic thin shell energy. We incorporate this basis in a functional map framework and demonstrate the effectiveness of our approach for mapping non-isometric shapes such that prominent features are put in correspondence. Finally, we describe the necessary adaptations to the functional map framework for working with non-orthogonal basis functions, thus considerably widening the scope of future uses of spectral shape correspondence.

Recently, there has been exciting progress in frame interpolation for rendered content. In this offline rendering setting, additional inputs, such as albedo and depth, can be extracted from a scene at a very low cost and, when integrated in a suitable fashion, can significantly improve the quality of the interpolated frames. Although existing approaches have been able to show good results, most high-quality interpolation methods use a synthesis network for direct color prediction. In complex scenarios, this can result in unpredictable behavior and lead to color artifacts. To mitigate this and to increase robustness, we propose to estimate the interpolated frame by predicting spatially varying kernels that operate on image splats. Kernel prediction ensures a linear mapping from the input images to the output and enables new opportunities, such as consistent and efficient interpolation of alpha values or many other additional channels and render passes that might exist. Additionally, we present an adaptive strategy that allows predicting full or partial keyframes that should be rendered with color samples solely based on the auxiliary features of a shot. This content-based spatio-temporal adaptivity allows rendering significantly fewer color pixels as compared to a fixed-step scheme when wanting to maintain a certain quality. Overall, these contributions lead to a more robust method and significant further reductions of the rendering costs.

Recent advancements in hardware-accelerated raytracing made it possible to achieve interactive framerates even for algorithms previously considered offline, such as path tracing. Interactive path tracing pipelines rely heavily on spatiotemporal denoising to produce a high-quality output from low-sample-count renderings. Such denoising is typically implemented as multiscale-kernel-based filters driven by lightweight U-Nets operating on pixels, and encoders operating on samples. In this work, we present a novel kernel architecture in the line of low-pass pyramid filters. Our architecture avoids the issues with the low-frequency response of previous such filters, resolving ringing, blotchiness, and box-shaped artefacts while improving overall detail. Instead of using classical downsampling and upsampling approaches, which are prone to aliasing, we let our weight predictor networks learn to partition the input radiance between pyramidal layers, predict kernels for denoising each partitioned and downscaled image, and then guide the upsampling process when combining layers. We present failure cases of pyramidal scale-composition in previous work and, through Fourier analysis, show how our method resolves them. Finally, we demonstrate state-of-the-art denoising performance.

We present Multi-feature Radiance-Predicting Neural Networks (MRPNN), a practical framework with a lightweight feature fusion neural network for rendering high-order scattered radiance of participating media in real time. By reformulating the Radiative Transfer Equation (RTE) through theoretical examination, we propose transmittance fields, generated at a low cost, as auxiliary information to help the network better approximate the RTE, drastically reducing the size of the neural network. The light weight network efficiently estimates the difficult-to-solve in-scattering term and allows for configurable shading parameters while improving prediction accuracy. In addition, we propose a frequency-sensitive stencil design in order to handle non-cloud shapes, resulting in accurate shadow boundaries. Results show that our MRPNN is able to synthesize indistinguishable output compared to the ground truth. Most importantly, MRPNN achieves a speedup of two orders of magnitude compared to the state-of-the-art, and is able to render high-quality participating material in real time.

Replicating a user’s pose from only wearable sensors is important for many AR/VR applications. Most existing methods for motion tracking avoid environment interaction apart from foot-floor contact due to their complex dynamics and hard constraints. However, in daily life people regularly interact with their environment, e.g. by sitting on a couch or leaning on a desk. Using Reinforcement Learning, we show that headset and controller pose, if combined with physics simulation and environment observations can generate realistic full-body poses even in highly constrained environments. The physics simulation automatically enforces the various constraints necessary for realistic poses, instead of manually specifying them as in many kinematic approaches. These hard constraints allow us to achieve high-quality interaction motions without typical artifacts such as penetration or contact sliding. We discuss three features, the environment representation, the contact reward and scene randomization, crucial to the performance of the method. We demonstrate the generality of the approach through various examples, such as sitting on chairs, a couch and boxes, stepping over boxes, rocking a chair and turning an office chair. We believe these are some of the highest-quality results achieved for motion tracking from sparse sensor with scene interaction.

Movement is how people interact with and affect their environment. For realistic character animation, it is necessary to synthesize such interactions between virtual characters and their surroundings. Despite recent progress in character animation using machine learning, most systems focus on controlling an agent’s movements in fairly simple and homogeneous environments, with limited interactions with other objects. Furthermore, many previous approaches that synthesize human-scene interactions require significant manual labeling of the training data. In contrast, we present a system that uses adversarial imitation learning and reinforcement learning to train physically-simulated characters that perform scene interaction tasks in a natural and life-like manner. Our method learns scene interaction behaviors from large unstructured motion datasets, without manual annotation of the motion data. These scene interactions are learned using an adversarial discriminator that evaluates the realism of a motion within the context of a scene. The key novelty involves conditioning both the discriminator and the policy networks on scene context. We demonstrate the effectiveness of our approach through three challenging scene interaction tasks: carrying, sitting, and lying down, which require coordination of a character’s movements in relation to objects in the environment. Our policies learn to seamlessly transition between different behaviors like idling, walking, and sitting. By randomizing the properties of the objects and their placements during training, our method is able to generalize beyond the objects and scenarios depicted in the training dataset, producing natural character-scene interactions for a wide variety of object shapes and placements. The approach takes physics-based character motion generation a step closer to broad applicability. Please see our supplementary video for more results.

We present a method to animate a character incorporating multiple part-wise motion priors (PMP). While previous works allow creating realistic articulated motions from reference data, the range of motion is largely limited by the available samples. Especially for the interaction-rich scenarios, it is impractical to attempt acquiring every possible interacting motion, as the combination of physical parameters increases exponentially. The proposed PMP allows us to assemble multiple part skills to animate a character, creating a diverse set of motions with different combinations of existing data. In our pipeline, we can train an agent with a wide range of part-wise priors. Therefore, each body part can obtain a kinematic insight of the style from the motion captures, or at the same time extract dynamics-related information from the additional part-specific simulation. For example, we can first train a general interaction skill, e.g. grasping, only for the dexterous part, and then combine the expert trajectories from the pre-trained agent with the kinematic priors of other limbs. Eventually, our whole-body agent learns a novel physical interaction skill even with the absence of the object trajectories in the reference motion sequence.

We present a method for reproducing complex multi-character interactions for physically simulated humanoid characters using deep reinforcement learning. Our method learns control policies for characters that imitate not only individual motions, but also the interactions between characters, while maintaining balance and matching the complexity of reference data. Our approach uses a novel reward formulation based on an interaction graph that measures distances between pairs of interaction landmarks. This reward encourages control policies to efficiently imitate the character’s motion while preserving the spatial relationships of the interactions in the reference motion. We evaluate our method on a variety of activities, from simple interactions such as a high-five greeting to more complex interactions such as gymnastic exercises, Salsa dancing, and box carrying and throwing. This approach can be used to “clean-up” existing motion capture data to produce physically plausible interactions or to retarget motion to new characters with different sizes, kinematics or morphologies while maintaining the interactions in the original data.

While progress has been made in the field of portrait reenactment, the problem of how to efficiently produce high-fidelity and accurate videos remains. Recent studies build direct mappings between driving signals and their predictions, leading to failure cases when synthesizing background textures and detailed local motions. In this paper, we propose the Video Portrait via Grid-based Codebook (VPGC) framework, which achieves efficient and high-fidelity portrait modeling. Our key insight is to query driving signals in a position-aware textural codebook with an explicit grid structure. The grid-based codebook stores delicate textural information locally according to our observations on video portraits, which can be learned efficiently and precisely. We subsequently design a Prior-Guided Driving Module to predict reliable features from the driving signals, which can be later decoded back to high-quality video portraits by querying the codebook. Comprehensive experiments are conducted to validate the effectiveness of our approach.

Face reenactment methods attempt to restore and re-animate portrait videos as realistically as possible. Existing methods face a dilemma in quality versus controllability: 2D GAN-based methods achieve higher image quality but suffer in fine-grained control of facial attributes compared with 3D counterparts. In this work, we propose StyleAvatar, a real-time photo-realistic portrait avatar reconstruction method using StyleGAN-based networks, which can generate high-fidelity portrait avatars with faithful expression control. We expand the capabilities of StyleGAN by introducing a compositional representation and a sliding window augmentation method, which enable faster convergence and improve translation generalization. Specifically, we divide the portrait scenes into three parts for adaptive adjustments: facial region, non-facial foreground region, and the background. Besides, our network leverages the best of UNet, StyleGAN and time coding for video learning, which enables high-quality video generation. Furthermore, a sliding window augmentation method together with a pre-training strategy are proposed to improve translation generalization and training performance, respectively. The proposed network can converge within two hours while ensuring high image quality and a forward rendering time of only 20 milliseconds. Furthermore, we propose a real-time live system, which further pushes research into applications. Results and experiments demonstrate the superiority of our method in terms of image quality, full portrait video generation, and real-time re-animation compared to existing facial reenactment methods. Training and inference code for this paper are at https://github.com/LizhenWangT/StyleAvatar.

We propose an end-to-end deep-learning approach for automatic rigging and retargeting of 3D models of human faces in the wild. Our approach, called Neural Face Rigging (NFR), holds three key properties: (i) NFR’s expression space maintains human-interpretable editing parameters for artistic controls; (ii) NFR is readily applicable to arbitrary facial meshes with different connectivity and expressions; (iii) NFR can encode and produce fine-grained details of complex expressions performed by arbitrary subjects. To the best of our knowledge, NFR is the first approach to provide realistic and controllable deformations of in-the-wild facial meshes, without the manual creation of blendshapes or correspondence. We design a deformation autoencoder and train it through a multi-dataset training scheme, which benefits from the unique advantages of two data sources: a linear 3DMM with interpretable control parameters as in FACS and 4D captures of real faces with fine-grained details. Through various experiments, we show NFR’s ability to automatically produce realistic and accurate facial deformations across a wide range of existing datasets and noisy facial scans in-the-wild, while providing artist-controlled, editable parameters.

Modern data-driven image generation models often surpass traditional graphics techniques in quality. However, while traditional modeling and animation tools allow precise control over the image generation process in terms of interpretable quantities — e.g., shapes and reflectances — endowing learned models with such controls is generally difficult.

In the context of human faces, we seek a data-driven generator architecture that simultaneously retains the photorealistic quality of modern generative adversarial networks (GAN) and allows explicit, disentangled controls over head shapes, expressions, identity, background, and illumination. While our high-level goal is shared by a large body of previous work, we approach the problem with a different philosophy: We treat the problem as an unconditional synthesis task, and engineer interpretable inductive biases into the model that make it easy for the desired behavior to emerge. Concretely, our generator is a combination of learned neural networks and fixed-function blocks, such as a 3D morphable head model and texture-mapping rasterizer, and we leave it up to the training process to figure out how they should be used together. This greatly simplifies the training problem by removing the need for labeled training data; we learn the distributions of the independent variables that drive the model instead of requiring that their values are known for each training image. Furthermore, we need no contrastive or imitation learning for correct behavior.

We show that our design successfully encourages the generative model to make use of the internal, interpretable representations in a semantically meaningful manner. This allows sampling of different aspects of the image independently, as well as precise control of the results by manipulating the internal state of the interpretable blocks within the generator. This enables, for instance, facial animation using traditional animation tools.

We propose ClipFace, a novel self-supervised approach for text-guided editing of textured 3D morphable model of faces. Specifically, we employ user-friendly language prompts to enable control of the expressions as well as appearance of 3D faces. We leverage the geometric expressiveness of 3D morphable models, which inherently possess limited controllability and texture expressivity, and develop a self-supervised generative model to jointly synthesize expressive, textured, and articulated faces in 3D. We enable high-quality texture generation for 3D faces by adversarial self-supervised training, guided by differentiable rendering against collections of real RGB images. Controllable editing and manipulation are given by language prompts to adapt texture and expression of the 3D morphable model. To this end, we propose a neural network that predicts both texture and expression latent codes of the morphable model. Our model is trained in a self-supervised fashion by exploiting differentiable rendering and losses based on a pre-trained CLIP model. Once trained, our model jointly predicts face textures in UV-space, along with expression parameters to capture both geometry and texture changes in facial expressions in a single forward pass. We further show the applicability of our method to generate temporally changing textures for a given animation sequence.

Neural radiance fields (NeRF) have achieved impressive performances in view synthesis by encoding neural representations of a scene. However, NeRFs require hundreds of images per scene to synthesize photo-realistic novel views. Training them on sparse input views leads to overfitting and incorrect scene depth estimation resulting in artifacts in the rendered novel views. Sparse input NeRFs were recently regularized by providing dense depth estimated from pre-trained networks as supervision, to achieve improved performance over sparse depth constraints. However, we find that such depth priors may be inaccurate due to generalization issues. Instead, we hypothesize that the visibility of pixels in different input views can be more reliably estimated to provide dense supervision. In this regard, we compute a visibility prior through the use of plane sweep volumes, which does not require any pre-training. By regularizing the NeRF training with the visibility prior, we successfully train the NeRF with few input views. We reformulate the NeRF to also directly output the visibility of a 3D point from a given viewpoint to reduce the training time with the visibility constraint. On multiple datasets, our model outperforms the competing sparse input NeRF models including those that use learned priors. The source code for our model can be found on our project page: https://nagabhushansn95.github.io/publications/2023/ViP-NeRF.html.

Neural Radiance Fields (NeRF) are a rapidly growing area of research with wide-ranging applications in computer vision, graphics, robotics, and more. In order to streamline the development and deployment of NeRF research, we propose a modular PyTorch framework, Nerfstudio. Our framework includes plug-and-play components for implementing NeRF-based methods, which make it easy for researchers and practitioners to incorporate NeRF into their projects. Additionally, the modular design enables support for extensive real-time visualization tools, streamlined pipelines for importing captured in-the-wild data, and tools for exporting to video, point cloud and mesh representations. The modularity of Nerfstudio enables the development of Nerfacto, our method that combines components from recent papers to achieve a balance between speed and quality, while also remaining flexible to future modifications. To promote community-driven development, all associated code and data are made publicly available with open-source licensing.

This paper presents a novel neural implicit radiance representation for free viewpoint relighting from a small set of unstructured photographs of an object lit by a moving point light source different from the view position. We express the shape as a signed distance function modeled by a multi layer perceptron. In contrast to prior relightable implicit neural representations, we do not disentangle the different light transport components, but model both the local and global light transport at each point by a second multi layer perceptron that, in addition, to density features, the current position, the normal (from the signed distance function), view direction, and light position, also takes shadow and highlight hints to aid the network in modeling the corresponding high frequency light transport effects. These hints are provided as a suggestion, and we leave it up to the network to decide how to incorporate these in the final relit result. We demonstrate and validate our neural implicit representation on synthetic and real scenes exhibiting a wide variety of shapes, material properties, and global illumination light transport.

Neural Radiance Fields (NeRF) have shown promising results in novel view synthesis. While achieving state-of-the-art rendering results, NeRF usually encodes all properties related to geometry and appearance of the scene together into several MLP (Multi-Layer Perceptron) networks, which hinders downstream manipulation of geometry, appearance and illumination. Recently researchers made attempts to edit geometry, appearance and lighting for NeRF. However, they fail to render view-consistent results after editing the appearance of the input scene. Moreover, high-frequency environmental relighting is also beyond their capability as lighting is modeled as Spherical Gaussian (SG) and Spherical Harmonic (SH) functions or a low-resolution environment map. To solve the above problems, we propose DE-NeRF to decouple view-independent appearance and view-dependent appearance in the scene with a hybrid lighting representation. Specifically, we first train a signed distance function to reconstruct an explicit mesh for the input scene. Then a decoupled NeRF learns to attach view-independent appearance to the reconstructed mesh by defining learnable disentangled features representing geometry and view-independent appearance on its vertices. For lighting, we approximate it with an explicit learnable environment map and an implicit lighting network to support both low-frequency and high-frequency relighting. By modifying the view-independent appearance, rendered results are consistent across different viewpoints. Our method also supports high-frequency environmental relighting by replacing the explicit environment map with a novel one and fitting the implicit lighting network to the novel environment map. Experiments show that our method achieves better editing and relighting performance both quantitatively and qualitatively compared to previous methods.

The problem of placing evenly-spaced stripes on a triangular mesh mirrors that of having evenly-spaced course rows and wale columns in a knit graph for a given geometry. This work presents strategies for producing helix-free stripe patterns and traces them to produce helix-free knit graphs suitable for machine knitting. We optimize directly for the discrete differential (1-form) of the stripe texture function, i.e., the spinning form, and demonstrate the knitting-specific advantages of this framework. In particular, we note how simple linear constraints allow us to place stitch irregularities, align course rows and wale columns to boundary/feature curves, and eliminate helical stripes. Two mixed-integer optimization strategies using these constraints are presented and applied to several mesh models. The results are smooth, globally-informed, helix-free stripe patterns that we trace to produce machine-knittable graphs. We further provide an explicit characterization of helical stripes and a theoretical analysis of their elimination constraints.

Sum-of-Squares Programming (SOSP) has recently been introduced to graphics as a unified way to address a large set of difficult problems involving higher order primitives. Unfortunately, a challenging aspect of this approach is the computational cost—especially for problems involving multiple geometries like collision detection. In this paper, we present techniques to reduce the cost of SOSP significantly. We use these improvements to speed up difficult problems like collision detection between Bézier triangles by as much as 300 ×. In addition, motivated by hair bundle simulation, we present SOSP based collision detection on the tapered cubic cylinder. We also present an algebraic formulation of rigid body motion enabling SOSP based collision detection for curved geometries and trajectories simultaneously. While these new formulations are complex, our speedups make them feasible. These advances improve the applicability of SOSP based collision detection and enable the continued progress of higher-order geometry processing.

High-order bases provide major advantages over linear ones in terms of efficiency, as they provide (for the same physical model) higher accuracy for the same running time, and reliability, as they are less affected by locking artifacts and mesh quality. Thus, we introduce a high-order finite element (FE) formulation (high-order bases) for elastodynamic simulation on high-order (curved) meshes with contact handling based on the recently proposed Incremental Potential Contact (IPC) model.

Our approach is based on the observation that each IPC optimization step used to minimize the elasticity, contact, and friction potentials leads to linear trajectories even in the presence of nonlinear meshes or nonlinear FE bases. It is thus possible to retain the strong non-penetration guarantees and large time steps of the original formulation while benefiting from the high-order bases and high-order geometry. We accomplish this by mapping displacements and resulting contact forces between a linear collision proxy and the underlying high-order representation.

We demonstrate the effectiveness of our approach in a selection of problems from graphics, computational fabrication, and scientific computing.

Synthesizing visual content that meets users’ needs often requires flexible and precise controllability of the pose, shape, expression, and layout of the generated objects. Existing approaches gain controllability of generative adversarial networks (GANs) via manually annotated training data or a prior 3D model, which often lack flexibility, precision, and generality. In this work, we study a powerful yet much less explored way of controlling GANs, that is, to "drag" any points of the image to precisely reach target points in a user-interactive manner, as shown in Fig.1. To achieve this, we propose DragGAN, which consists of two main components: 1) a feature-based motion supervision that drives the handle point to move towards the target position, and 2) a new point tracking approach that leverages the discriminative generator features to keep localizing the position of the handle points. Through DragGAN, anyone can deform an image with precise control over where pixels go, thus manipulating the pose, shape, expression, and layout of diverse categories such as animals, cars, humans, landscapes, etc. As these manipulations are performed on the learned generative image manifold of a GAN, they tend to produce realistic outputs even for challenging scenarios such as hallucinating occluded content and deforming shapes that consistently follow the object’s rigidity. Both qualitative and quantitative comparisons demonstrate the advantage of DragGAN over prior approaches in the tasks of image manipulation and point tracking. We also showcase the manipulation of real images through GAN inversion.

In this paper we present Diffusion Image Analogies—an example-based image editing approach that builds upon the concept of image analogies originally introduced by Hertzmann et al. [2001]. Given a pair of images that specify the intent of a specific transition, our approach enables to modify the target image in a way that it follows the analogy specified by this exemplar. In contrast to previous techniques which were able to capture analogies mostly on the low-level textural details our approach handles also changes in higher level semantics including transition of object domain, change of facial expression, or stylization. Although similar modifications can be achieved using diffusion models guided by text prompts [Rombach et al. 2022] our approach can operate solely in the domain of images without the need to specify the user’s intent using textual form. We demonstrate power of our approach in various challenging scenarios where the specified analogy would be difficult to transfer using previous techniques.

This paper presents a novel neural material relighting method for revisualizing a photograph of a planar spatially-varying material under novel viewing and lighting conditions. Our approach is motivated by the observation that the plausibility of a spatially varying material is judged purely on the visual appearance, not on the underlying distribution of appearance parameters. Therefore, instead of using an intermediate parametric representation (e.g., SVBRDF) that requires a rendering stage to visualize the spatially-varying material for novel viewing and lighting conditions, neural material relighting directly generates the target visual appearance. We explore and evaluate two different use cases where the relit results are either used directly, or where the relit images are used to enhance the input in existing multi-image spatially varying reflectance estimation methods. We demonstrate the robustness and efficacy for both use cases on a wide variety of spatially varying materials.

Bidirectional Texture Functions (BTFs) are able to represent complex materials with greater generality than traditional analytical models. This holds true for both measured real materials and synthetic ones. Recent advancements in neural BTF representations have significantly reduced storage costs, making them more practical for use in rendering. These representations typically combine spatial feature (latent) textures with neural decoders that handle angular dimensions per spatial location. However, these models have yet to combine fast compression and inference, accuracy, and generality. In this paper, we propose a biplane representation for BTFs, which uses a feature texture in the half-vector domain as well as the spatial domain. This allows the learned representation to encode high-frequency details in both the spatial and angular domains. Our decoder is small yet general, meaning it is trained once and fixed. Additionally, we optionally combine this representation with a neural offset module for parallax and masking effects. Our model can represent a broad range of BTFs and has fast compression and inference due to its lightweight architecture. Furthermore, it enables a simple way to capture BTF data. By taking about 20 cell phone photos with a collocated camera and flash, our model can plausibly recover the entire BTF, despite never observing function values with differing view and light directions. We demonstrate the effectiveness of our model in the acquisition of many measured materials, including challenging materials such as fabrics.

We present a technique for automatically producing a deformation of an input triangle mesh, guided solely by a text prompt. Our framework is capable of deformations that produce both large, low-frequency shape changes, and small high-frequency details. Our framework relies on differentiable rendering to connect geometry to powerful pre-trained image encoders, such as CLIP and DINO. Notably, updating mesh geometry by taking gradient steps through differentiable rendering is notoriously challenging, commonly resulting in deformed meshes with significant artifacts. These difficulties are amplified by noisy and inconsistent gradients from CLIP. To overcome this limitation, we opt to represent our mesh deformation through Jacobians, which updates deformations in a global, smooth manner (rather than locally-sub-optimal steps). Our key observation is that Jacobians are a representation that favors smoother, large deformations, leading to a global relation between vertices and pixels, and avoiding localized noisy gradients. Additionally, to ensure the resulting shape is coherent from all 3D viewpoints, we encourage the deep features computed on the 2D encoding of the rendering to be consistent for a given vertex from all viewpoints. We demonstrate that our method is capable of smoothly-deforming a wide variety of source mesh and target text prompts, achieving both large modifications to, e.g., body proportions of animals, as well as adding fine semantic details, such as shoe laces on an army boot and fine details of a face.

There is a growing demand for the accessible creation of high-quality 3D avatars that are animatable and customizable. Although 3D morphable models provide intuitive control for editing and animation, and robustness for single-view face reconstruction, they cannot easily capture geometric and appearance details. Methods based on neural implicit representations, such as signed distance functions (SDF) or neural radiance fields, approach photo-realism, but are difficult to animate and do not generalize well to unseen data. To tackle this problem, we propose a novel method for constructing implicit 3D morphable face models that are both generalizable and intuitive for editing. Trained from a collection of high-quality 3D scans, our face model is parameterized by geometry, expression, and texture latent codes with a learned SDF and explicit UV texture parameterization. Once trained, we can reconstruct an avatar from a single in-the-wild image by leveraging the learned prior to project the image into the latent space of our model. Our implicit morphable face models can be used to render an avatar from novel views, animate facial expressions by modifying expression codes, and edit textures by directly painting on the learned UV-texture maps. We demonstrate quantitatively and qualitatively that our method improves upon photo-realism, geometry, and expression accuracy compared to state-of-the-art methods.

The recent proliferation of 3D content that can be consumed on hand-held devices necessitates efficient tools for transmitting large geometric data, e.g., 3D meshes, over the Internet. Detailed high-resolution assets can pose a challenge to storage as well as transmission bandwidth, and level-of-detail techniques are often used to transmit an asset using an appropriate bandwidth budget. It is especially desirable for these methods to transmit data progressively, improving the quality of the geometry with more data. Our key insight is that the geometric details of 3D meshes often exhibit similar local patterns even across different shapes, and thus can be effectively represented with a shared learned generative space. We learn this space using a subdivision-based encoder-decoder architecture trained in advance on a large collection of surfaces. We further observe that additional residual features can be transmitted progressively between intermediate levels of subdivision that enable the client to control the tradeoff between bandwidth cost and quality of reconstruction, providing a neural progressive mesh representation. We evaluate our method on a diverse set of complex 3D shapes and demonstrate that it outperforms baselines in terms of compression ratio and reconstruction quality.

3D facial avatar reconstruction has been a significant research topic in computer graphics and computer vision, where photo-realistic rendering and flexible controls over poses and expressions are necessary for many related applications. Recently, its performance has been greatly improved with the development of neural radiance fields (NeRF). However, most existing NeRF-based facial avatars focus on subject-specific reconstruction and reenactment, requiring multi-shot images containing different views of the specific subject for training, and the learned model cannot generalize to new identities, limiting its further applications. In this work, we propose a one-shot 3D facial avatar reconstruction framework that only requires a single source image to reconstruct a high-fidelity 3D facial avatar. For the challenges of lacking generalization ability and missing multi-view information, we leverage the generative prior of 3D GAN and develop an efficient encoder-decoder network to reconstruct the canonical neural volume of the source image, and further propose a compensation network to complement facial details. To enable fine-grained control over facial dynamics, we propose a deformation field to warp the canonical volume into driven expressions. Through extensive experimental comparisons, we achieve superior synthesis results compared to several state-of-the-art methods.

Existing approaches to animatable NeRF-based head avatars are either built upon face templates or use the expression coefficients of templates as the driving signal. Despite the promising progress, their performances are heavily bound by the expression power and the tracking accuracy of the templates. In this work, we present LatentAvatar, an expressive neural head avatar driven by latent expression codes. Such latent expression codes are learned in an end-to-end and self-supervised manner without templates, enabling our method to get rid of expression and tracking issues. To achieve this, we leverage a latent head NeRF to learn the person-specific latent expression codes from a monocular portrait video, and further design a Y-shaped network to learn the shared latent expression codes of different subjects for cross-identity reenactment. By optimizing the photometric reconstruction objectives in NeRF, the latent expression codes are learned to be 3D-aware while faithfully capturing the high-frequency detailed expressions. Moreover, by learning a mapping between the latent expression code learned in shared and person-specific settings, LatentAvatar is able to perform expressive reenactment between different subjects. Experimental results show that our LatentAvatar is able to capture challenging expressions and the subtle movement of teeth and even eyeballs, which outperforms previous state-of-the-art solutions in both quantitative and qualitative comparisons. Project page: https://www.liuyebin.com/latentavatar.