Diffusion models currently dominate the field of data- driven image synthesis with their unparalleled scaling to large datasets. In this paper, we identify and rectify several causes for uneven and ineffective training in the popular ADM diffusion model architecture, without altering its high- level structure. Observing uncontrolled magnitude changes and imbalances in both the network activations and weights over the course of training, we redesign the network layers to preserve activation, weight, and update magnitudes on ex- pectation. We find that systematic application of this philoso- phy eliminates the observed drifts and imbalances, resulting in considerably better networks at equal computational com- plexity. Our modifications improve the previous record FID of 2.41 in ImageNet-512 synthesis to 1.81, achieved using fast deterministic sampling. As an independent contribution, we present a method for setting the exponential moving average (EMA) parameters post-hoc, i.e., after completing the training run. This allows precise tuning of EMA length without the cost of performing several training runs, and reveals its surprising interactions with network architecture, training time, and guidance.

The primary axes of interest in image-generating diffusion models are image quality, the amount of variation in the results, and how well the results align with a given condition, e.g., a class label or a text prompt. The popular classifier-free guidance approach uses an unconditional model to guide a conditional model, leading to simultaneously better prompt alignment and higher-quality images at the cost of reduced variation. These effects seem inherently entangled, and thus hard to control. We make the surprising observation that it is possible to obtain disentangled control over image quality without compromising the amount of variation by guiding generation using a smaller, less-trained version of the model itself rather than an unconditional model. This leads to significant improvements in ImageNet generation, setting record FIDs of 1.01 for 64x64 and 1.25 for 512x512, using publicly available networks. Furthermore, the method is also applicable to unconditional diffusion models, drastically improving their quality.

Guidance is a crucial technique for extracting the best performance out of image-generating diffusion models. Traditionally, a constant guidance weight has been applied throughout the sampling chain of an image. We show that guidance is clearly harmful toward the beginning of the chain (high noise levels), largely unnecessary toward the end (low noise levels), and only beneficial in the middle. We thus restrict it to a specific range of noise levels, improving both the inference speed and result quality. This limited guidance interval improves the record FID in ImageNet-512 significantly, from 1.81 to 1.40. We show that it is quantitatively and qualitatively beneficial across different sampler parameters, network architectures, and datasets, including the large-scale setting of Stable Diffusion XL. We thus suggest exposing the guidance interval as a hyperparameter in all diffusion models that use guidance.

Text-to-image synthesis has recently seen significant progress thanks to large pretrained language models, large-scale training data, and the introduction of scalable model families such as diffusion and autoregressive models. However, the best-performing models require iterative evaluation to generate a single sample. In contrast, generative adversarial networks (GANs) only need a single forward pass. They are thus much faster, but they currently remain far behind the state-of-the-art in large-scale text-to-image synthesis. This paper aims to identify the necessary steps to regain competitiveness. Our proposed model, StyleGAN-T, addresses the specific requirements of large-scale text-to-image synthesis, such as large capacity, stable training on diverse datasets, strong text alignment, and controllable variation vs. text alignment tradeoff. StyleGAN-T significantly improves over previous GANs and outperforms distilled diffusion models - the previous state-of-the-art in fast text-to-image synthesis - in terms of sample quality and speed.

We present a video generation model that accurately reproduces object motion, changes in camera viewpoint, and new content that arises over time. Existing video generation methods often fail to produce new content as a function of time while maintaining consistencies expected in real environments, such as plausible dynamics and object persistence. A common failure case is for content to never change due to over-reliance on inductive bias to provide temporal consistency, such as a single latent code that dictates content for the entire video. On the other extreme, without long-term consistency, generated videos may morph unrealistically between different scenes. To address these limitations, we prioritize the time axis by redesigning the temporal latent representation and learning long-term consistency from data by training on longer videos. We leverage a two-phase training strategy, where we separately train using longer videos at a low resolution and shorter videos at a high resolution. To evaluate the capabilities of our model, we introduce two new benchmark datasets with explicit focus on long-term temporal dynamics.

We argue that the theory and practice of diffusion-based generative models are currently unnecessarily convoluted and seek to remedy the situation by presenting a design space that clearly separates the concrete design choices. This lets us identify several changes to both the sampling and training processes, as well as preconditioning of the score networks. Together, our improvements yield new state-of-the-art FID of 1.79 for CIFAR-10 in a class-conditional setting and 1.97 in an unconditional setting, with much faster sampling (35 network evaluations per image) than prior designs. To further demonstrate their modular nature, we show that our design changes dramatically improve both the efficiency and quality obtainable with pre-trained score networks from previous work, including improving the FID of a previously trained ImageNet-64 model from 2.07 to near-SOTA 1.55, and after re-training with our proposed improvements to a new SOTA of 1.36.

Time-lapse image sequences offer visually compelling insights into dynamic processes that are too slow to observe in real time. However, playing a long time-lapse sequence back as a video often results in distracting flicker due to random effects, such as weather, as well as cyclic effects, such as the day-night cycle. We introduce the problem of disentangling time-lapse sequences in a way that allows separate, after-the-fact control of overall trends, cyclic effects, and random effects in the images, and describe a technique based on data-driven generative models that achieves this goal. This enables us to "re-render" the sequences in ways that would not be possible with the input images alone. For example, we can stabilize a long sequence to focus on plant growth over many months, under selectable, consistent weather. Our approach is based on Generative Adversarial Networks (GAN) that are conditioned with the time coordinate of the time-lapse sequence. Our architecture and training procedure are designed so that the networks learn to model random variations, such as weather, using the GAN's latent space, and to disentangle overall trends and cyclic variations by feeding the conditioning time label to the model using Fourier features with specific frequencies. We show that our models are robust to defects in the training data, enabling us to amend some of the practical difficulties in capturing long time-lapse sequences, such as temporary occlusions, uneven frame spacing, and missing frames.

We observe that despite their hierarchical convolutional nature, the synthesis process of typical generative adversarial networks depends on absolute pixel coordinates in an unhealthy manner. This manifests itself as, e.g., detail appearing to be glued to image coordinates instead of the surfaces of depicted objects. We trace the root cause to careless signal processing that causes aliasing in the generator network. Interpreting all signals in the network as continuous, we derive generally applicable, small architectural changes that guarantee that unwanted information cannot leak into the hierarchical synthesis process. The resulting networks match the FID of StyleGAN2 but differ dramatically in their internal representations, and they are fully equivariant to translation and rotation even at subpixel scales. Our results pave the way for generative models better suited for video and animation.

The style-based GAN architecture (StyleGAN) yields state-of-the-art results in data-driven unconditional generative image modeling. We expose and analyze several of its characteristic artifacts, and propose changes in both model architecture and training methods to address them. In particular, we redesign the generator normalization, revisit progressive growing, and regularize the generator to encourage good conditioning in the mapping from latent codes to images. In addition to improving image quality, this path length regularizer yields the additional benefit that the generator becomes significantly easier to invert. This makes it possible to reliably attribute a generated image to a particular network. We furthermore visualize how well the generator utilizes its output resolution, and identify a capacity problem, motivating us to train larger models for additional quality improvements. Overall, our improved model redefines the state of the art in unconditional image modeling, both in terms of existing distribution quality metrics as well as perceived image quality.

Training generative adversarial networks (GAN) using too little data typically leads to discriminator overfitting, causing training to diverge. We propose an adaptive discriminator augmentation mechanism that significantly stabilizes training in limited data regimes. The approach does not require changes to loss functions or network architectures, and is applicable both when training from scratch and when fine-tuning an existing GAN on another dataset. We demonstrate, on several datasets, that good results are now possible using only a few thousand training images, often matching StyleGAN2 results with an order of magnitude fewer images. We expect this to open up new application domains for GANs. We also find that the widely used CIFAR-10 is, in fact, a limited data benchmark, and improve the record FID from 5.59 to 2.42.

We present a modular differentiable renderer design that yields performance superior to previous methods by leveraging existing, highly optimized hardware graphics pipelines. Our design supports all crucial operations in a modern graphics pipeline: rasterizing large numbers of triangles, attribute interpolation, filtered texture lookups, as well as user-programmable shading and geometry processing, all in high resolutions. Our modular primitives allow custom, high-performance graphics pipelines to be built directly within automatic differentiation frameworks such as PyTorch or TensorFlow. As a motivating application, we formulate facial performance capture as an inverse rendering problem and show that it can be solved efficiently using our tools. Our results indicate that this simple and straightforward approach achieves excellent geometric correspondence between rendered results and reference imagery.

We propose an alternative generator architecture for generative adversarial networks, borrowing from style transfer literature. The new architecture leads to an automatically learned, unsupervised separation of high-level attributes (e.g., pose and identity when trained on human faces) and stochastic variation in the generated images (e.g., freckles, hair), and it enables intuitive, scale-specific control of the synthesis. The new generator improves the state-of-the-art in terms of traditional distribution quality metrics, leads to demonstrably better interpolation properties, and also better disentangles the latent factors of variation. To quantify interpolation quality and disentanglement, we propose two new, automated methods that are applicable to any generator architecture. Finally, we introduce a new, highly varied and high-quality dataset of human faces.

Unsupervised image-to-image translation methods learn to map images in a given class to an analogous image in a different class, drawing on unstructured (non-registered) datasets of images. While remarkably successful, current methods require access to many images in both source and destination classes at training time. We argue this greatly limits their use. Drawing inspiration from the human capability of picking up the essence of a novel object from a small number of examples and generalizing from there, we seek a few-shot, unsupervised image-to-image translation algorithm that works on previously unseen target classes that are specified, at test time, only by a few example images. Our model achieves this few-shot generation capability by coupling an adversarial training scheme with a novel network design. Through extensive experimental validation and comparisons to several baseline methods on benchmark datasets, we verify the effectiveness of the proposed framework. Our implementation and datasets are available at https://github.com/NVlabs/FUNIT.

The ability to automatically estimate the quality and coverage of the samples produced by a generative model is a vital requirement for driving algorithm research. We present an evaluation metric that can separately and reliably measure both of these aspects in image generation tasks by forming explicit, non-parametric representations of the manifolds of real and generated data. We demonstrate the effectiveness of our metric in StyleGAN and BigGAN by providing several illustrative examples where existing metrics yield uninformative or contradictory results. Furthermore, we analyze multiple design variants of StyleGAN to better understand the relationships between the model architecture, training methods, and the properties of the resulting sample distribution. In the process, we identify new variants that improve the state-of-the-art. We also perform the first principled analysis of truncation methods and identify an improved method. Finally, we extend our metric to estimate the perceptual quality of individual samples, and use this to study latent space interpolations.

We describe a novel method for training high-quality image denoising models based on unorganized collections of corrupted images. The training does not need access to clean reference images, or explicit pairs of corrupted images, and can thus be applied in situations where such data is unacceptably expensive or impossible to acquire. We build on a recent technique that removes the need for reference data by employing networks with a "blind spot" in the receptive field, and significantly improve two key aspects: image quality and training efficiency. Our result quality is on par with state-of-the-art neural network denoisers in the case of i.i.d. additive Gaussian noise, and not far behind with Poisson and impulse noise. We also successfully handle cases where parameters of the noise model are variable and/or unknown in both training and evaluation data.

We apply basic statistical reasoning to signal reconstruction by machine learning - learning to map corrupted observations to clean signals - with a simple and powerful conclusion: it is possible to learn to restore images by only looking at corrupted examples, at performance at and sometimes exceeding training using clean data, without explicit image priors or likelihood models of the corruption. In practice, we show that a single model learns photographic noise removal, denoising synthetic Monte Carlo images, and reconstruction of undersampled MRI scans - all corrupted by different processes - based on noisy data only.

We describe a machine learning technique for reconstructing image sequences rendered using Monte Carlo methods. Our primary focus is on reconstruction of global illumination with extremely low sampling budgets at interactive rates. Motivated by recent advances in image restoration with deep convolutional networks, we propose a variant of these networks better suited to the class of noise present in Monte Carlo rendering. We allow for much larger pixel neighborhoods to be taken into account, while also improving execution speed by an order of magnitude. Our primary contribution is the addition of recurrent connections to the network in order to drastically improve temporal stability for sequences of sparsely sampled input images. Our method also has the desirable property of automatically modeling relationships based on auxiliary per-pixel input channels, such as depth and normals. We show significantly higher quality results compared to existing methods that run at comparable speeds, and furthermore argue a clear path for making our method run at realtime rates in the near future.

We present a machine learning technique for driving 3D facial animation by audio input in real time and with low latency. Our deep neural network learns a mapping from input waveforms to the 3D vertex coordinates of a face model, and simultaneously discovers a compact, latent code that disambiguates the variations in facial expression that cannot be explained by the audio alone. During inference, the latent code can be used as an intuitive control for the emotional state of the face puppet. We train our network with 3--5 minutes of high-quality animation data obtained using traditional, vision-based performance capture methods. Even though our primary goal is to model the speaking style of a single actor, our model yields reasonable results even when driven with audio from other speakers with different gender, accent, or language, as we demonstrate with a user study. The results are applicable to in-game dialogue, low-cost localization, virtual reality avatars, and telepresence.

We extend parametric texture synthesis to capture rich, spatially varying parametric reflectance models from a single image. Our input is a single head-lit flash image of a mostly flat, mostly stationary (textured) surface, and the output is a tile of SVBRDF parameters that reproduce the appearance of the material. No user intervention is required. Our key insight is to make use of a recent, powerful texture descriptor based on deep convolutional neural network statistics for "softly" comparing the model prediction and the examplars without requiring an explicit point-to-point correspondence between them. This is in contrast to traditional reflectance capture that requires pointwise constraints between inputs and outputs under varying viewing and lighting conditions. Seen through this lens, our method is an indirect algorithm for fitting photorealistic SVBRDFs. The problem is severely ill-posed and non-convex. To guide the optimizer towards desirable solutions, we introduce a soft Fourier-domain prior for encouraging spatial stationarity of the reflectance parameters and their correlations, and a complementary preconditioning technique that enables efficient exploration of such solutions by L-BFGS, a standard non-linear numerical optimizer.

We introduce a novel Metropolis rendering algorithm that directly computes image gradients, and reconstructs the final image from the gradients by solving a Poisson equation. The reconstruction is aided by a low-fidelity approximation of the image computed during gradient sampling. As an extension of path-space Metropolis light transport, our algorithm is well suited for difficult transport scenarios. We demonstrate that our method outperforms the state-of-the-art in several well-known test scenes. Additionally, we analyze the spectral properties of gradient-domain sampling, and compare it to the traditional image-domain sampling.

Stochastic techniques for rendering indirect illumination suffer from noise due to the variance in the integrand. In this paper, we describe a general reconstruction technique that exploits anisotropy in the light field and permits efficient reuse of input samples between pixels or world-space locations, multiplying the effective sampling rate by a large factor. Our technique introduces visibility-aware anisotropic reconstruction to indirect illumination, ambient occlusion and glossy reflections. It operates on point samples without knowledge of the scene, and can thus be seen as an advanced image filter. Our results show dramatic improvement in image quality while using very sparse input samplings.

Stochastic renderers produce unbiased but noisy images of scenes that include the advanced camera effects of motion and defocus blur and possibly other effects such as transparency. We present a simple algorithm that selectively adds bias in the form of image space blur to pixels that are unlikely to have high frequency content in the final image. For each pixel, we sweep once through a fixed neighborhood of samples in front to back order, using a simple accumulation scheme. We achieve good quality images with only 16 samples per pixel, making the algorithm potentially practical for interactive stochastic rendering in the near future.

We present a novel method for increasing the efficiency of stochastic rasterization of motion and defocus blur. Contrary to earlier approaches, our method is efficient even with the low sampling densities commonly encountered in realtime rendering, while allowing the use of arbitrary sampling patterns for maximal image quality. Our clipless dual-space formulation avoids problems with triangles that cross the camera plane during the shutter interval. The method is also simple to plug into existing rendering systems.

Traditionally, effects that require evaluating multidimensional integrals for each pixel, such as motion blur, depth of field, and soft shadows, suffer from noise due to the variance of the high-dimensional integrand. In this paper, we describe a general reconstruction technique that exploits the anisotropy in the temporal light field and permits efficient reuse of samples between pixels, multiplying the effective sampling rate by a large factor. We show that our technique can be applied in situations that are challenging or impossible for previous anisotropic reconstruction methods, and that it can yield good results with very sparse inputs. We demonstrate our method for simultaneous motion blur, depth of field, and soft shadows.

We describe the architecture of a novel system for precomputing sparse directional occlusion caches. These caches are used for accelerating a fast cinematic lighting pipeline that works in the spherical harmonics domain. The system was used as a primary lighting technology in the movie Avatar, and is able to efficiently handle massive scenes of unprecedented complexity through the use of a flexible, stream-based geometry processing architecture, a novel out-of-core algorithm for creating efficient ray tracing acceleration structures, and a novel out-of-core GPU ray tracing algorithm for the computation of directional occlusion and spherical integrals at arbitrary points.

We introduce a meshless hierarchical representation for solving light transport problems. Precomputed radiance transfer (PRT) and finite elements require a discrete representation of illumination over the scene. Non-hierarchical approaches such as per-vertex values are simple to implement, but lead to long precomputation. Hierarchical bases like wavelets lead to dramatic acceleration, but in their basic form they work well only on flat or smooth surfaces. We introduce a hierarchical function basis induced by scattered data approximation. It is decoupled from the geometric representation, allowing the hierarchical representation of illumination on complex objects. We present simple data structures and algorithms for constructing and evaluating the basis functions. Due to its hierarchical nature, our representation adapts to the complexity of the illumination, and can be queried at different scales. We demonstrate the power of the new basis in a novel precomputed direct-to-indirect light transport algorithm that greatly increases the complexity of scenes that can be handled by PRT approaches.

We present a novel architecture for hardware-accelerated rendering of point primitives. Our pipeline implements a refined version of EWA splatting, a high quality method for antialiased rendering of point sampled representations. A central feature of our design is the seamless integration of the architecture into conventional, OpenGL-like graphics pipelines so as to complement triangle-based rendering. The specific properties of the EWA algorithm required a variety of novel design concepts including a ternary depth test and using an on-chip pipelined heap data structure for making the memory accesses of splat primitives more coherent. In addition, we developed a computationally stable evaluation scheme for perspectively corrected splats. We implemented our architecture both on reconfigurable FPGA boards and as an ASIC prototype, and we integrated it into an OpenGL-like software implementation. Our evaluation comprises a detailed performance analysis using scenes of varying complexity.

Abstract Displaying a synthetic image on a computer display requires determining the colors of individual pixels. To avoid aliasing, multiple samples of the image can be taken per pixel, after which the color of a pixel may be computed as a weighted sum of the samples. The positions and weights of the samples play a major role in the resulting image quality, especially in real‐time applications where usually only a handful of samples can be afforded per pixel. This paper presents a new error metric and an optimization method for antialiasing patterns used in image reconstruction. The metric is based on comparing the pattern against a given reference reconstruction filter in spatial domain, and it takes into account psychovisually measured angle‐specific acuities for sharp features.

Abstract We identify and analyze several performance problems in a state‐of‐the‐art physically‐based soft shadow volume algorithm, and present an improved method that alleviates these problems by replacing an overly conservative spatial acceleration structure by a more efficient one. The new technique consistently outperforms both the previous method and a ray tracing‐based reference solution in several realistic situations while retaining the correctness of the solution and other desirable characteristics of the previous method. These include the unintrusiveness of the original algorithm, meaning that our method can be used as a black‐box shadow solver in any offline renderer without requiring multiple passes over the image or other special accommodation. We achieve speedup factors from 1.6 to 12.3 when compared to the previous method. Categories and Subject Descriptors (according to ACM CCS): I.3.7 [Three‐Dimensional Graphics and Realism]: Shadowing

We present a novel algorithm for rendering physically-based soft shadows in complex scenes.Instead of casting shadow rays, we place both the points to be shaded and the samples of an area light source into separate hierarchies, and compute hierarchically the shadows caused by each occluding triangle.This yields an efficient algorithm with memory requirements independent of the complexity of the scene.

We present a new, fast algorithm for rendering physically-based soft shadows in ray tracing-based renderers. Our method replaces the hundreds of shadow rays commonly used in stochastic ray tracers with a single shadow ray and a local reconstruction of the visibility function. Compared to tracing the shadow rays. our algorithm produces exactly the same image while executing one to two orders of magnitude faster in the test scenes used. Our first contribution is a two-stage method for quickly determining the silhouette edges that overlap an area light source, as seen from the point to be shaded. Secondly, we show that these partial silhouettes of occluders, along with a single shadow ray, are sufficient for reconstructing the visibility function between the point and the light source.

In causal processes decisions do not depend on future data. Many well-known problems, such as occlusion culling, order-independent transparency and edge antialiasing cannot be properly solved using the traditional causal rendering architectures, because future data may change the interpretation of current events.We propose adding a delay stream between the vertex and pixel processing units. While a triangle resides in the delay stream, subsequent triangles generate occlusion information. As a result, the triangle may be culled by primitives that were submitted after it. We show two-to fourfold efficiency improvements in pixel processing and video memory bandwidth usage in common benchmark scenes. We also demonstrate how the memory requirements of order-independent transparency can be substantially reduced by using delay streams. Finally, we describe how discontinuity edges can be detected in hardware. Previously used heuristics for collapsing samples in adaptive supersampling are thus replaced by connectivity information.