Both text-to-image generation and large language models (LLMs) have made significant advancements. However, many text-to-image models still employ the somewhat outdated T5 and CLIP as their text encoders. In this work, we investigate the effectiveness of using modern decoder-only LLMs as text encoders for text-to-image diffusion models. We build a standardized training and evaluation pipeline that allows us to isolate and evaluate the effect of different text embeddings. We train a total of 27 text-to-image models with 12 different text encoders to analyze the critical aspects of LLMs that could impact text-to-image generation, including the approaches to extract embeddings, different LLMs variants, and model sizes. Our experiments reveal that the de facto way of using last-layer embeddings as conditioning leads to inferior performance. Instead, we explore embeddings from various layers and find that using layer-normalized averaging across all layers significantly improves alignment with complex prompts. Most LLMs with this conditioning outperform the baseline T5 model, showing enhanced performance in advanced visio-linguistic reasoning skills.

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.

We present a diffusion-based model for 3D-aware generative novel view synthesis from as few as a single input image. Our model samples from the distribution of possible renderings consistent with the input and, even in the presence of ambiguity, is capable of rendering diverse and plausible novel views. To achieve this, our method makes use of existing 2D diffusion backbones but, crucially, incorporates geometry priors in the form of a 3D feature volume. This latent feature field captures the distribution over possible scene representations and improves our method’s ability to generate view-consistent novel renderings. In addition to generating novel views, our method has the ability to autoregressively synthesize 3D-consistent sequences. We demonstrate state-of-the-art results on synthetic renderings and room-scale scenes; we also show compelling results for challenging, real-world objects.

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.

Unsupervised generation of high-quality multi-view-consistent images and 3D shapes using only collections of single-view 2D photographs has been a long-standing challenge. Existing 3D GANs are either compute intensive or make approximations that are not 3D-consistent; the former limits quality and resolution of the generated images and the latter adversely affects multi-view consistency and shape quality. In this work, we improve the computational efficiency and image quality of 3D GANs without overly relying on these approximations. We introduce an expressive hybrid explicit implicit network architecture that, together with other design choices, synthesizes not only high-resolution multi-view-consistent images in real time but also produces high-quality 3D geometry. By decoupling feature generation and neural rendering, our framework is able to leverage state-of-the-art 2D CNN generators, such as StyleGAN2, and inherit their efficiency and expressiveness. We demonstrate state-of-the-art 3D-aware synthesis with FFHQ and AFHQ Cats, among other experiments.

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.

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.

Disentanglement learning is crucial for obtaining disentangled representations and controllable generation. Current disentanglement methods face several inherent limitations: difficulty with high-resolution images, primarily focusing on learning disentangled representations, and non-identifiability due to the unsupervised setting. To alleviate these limitations, we design new architectures and loss functions based on StyleGAN (Karras et al., 2019), for semi-supervised high-resolution disentanglement learning. We create two complex high-resolution synthetic datasets for systematic testing. We investigate the impact of limited supervision and find that using only 0.25% 2.5% of labeled data is sufficient for good disentanglement on both synthetic and real datasets. We propose new metrics to quantify generator controllability, and observe there may exist a crucial trade-off between disentangled representation learning and controllable generation. We also consider semantic fine-grained image editing to achieve better generalization to unseen images.

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 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.

Instant Radiosity and its derivatives are interactive methods for efficiently estimating global (indirect) illumination. They represent the last indirect bounce of illumination before the camera as the composite radiance field emitted by a set of virtual point light sources (VPLs). In complex scenes, current algorithms suffer from a difficult combination of two issues: it remains a challenge to distribute VPLs in a manner that simultaneously gives a high-quality indirect illumination solution for each frame, and to do so in a temporally coherent manner. We address both issues by building, and maintaining over time, an adaptive and temporally coherent distribution of VPLs in locations where they bring indirect light to the image. We introduce a novel heuristic sampling method that strives to only move as few of the VPLs between frames as possible. The result is, to the best of our knowledge, the first interactive global illumination algorithm that works in complex, highly-occluded scenes, suffers little from temporal flickering, supports moving cameras and light sources, and is output-sensitive in the sense that it places VPLs in locations that matter most to the final result.

Instant Radiosity and its derivatives are interactive methods for efficiently estimating global (indirect) illumination. They represent the last indirect bounce of illumination before the camera as the composite radiance field emitted by a set of virtual point light sources (VPLs). In complex scenes, current algorithms suffer from a difficult combination of two issues: it remains a challenge to distribute VPLs in a manner that simultaneously gives a high-quality indirect illumination solution for each frame, and does so in a temporally coherent manner. We address both issues by building, and maintaining over time, an adaptive and temporally coherent distribution of VPLs in locations where they bring indirect light to the image. We introduce a novel heuristic sampling method that strives to only move as few of the VPLs between frames as possible. The result is, to the best of our knowledge, the first interactive global illumination algorithm that works in complex, highly-occluded scenes, suffers little from temporal flickering, supports moving cameras and light sources, and is output-sensitive in the sense that it places VPLs in locations that matter most to the final result.

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.

Abstract The traditional method of rendering semi‐transparent surfaces using alpha blending requires sorting the surfaces in depth order. There are several techniques for order‐independent transparency, but most require either unbounded storage or can be fragile due to forced compaction of information during rendering. Stochastic transparency works in a fixed amount of storage and produces results with the correct expected value. However, carelessly chosen sampling strategies easily result in high variance of the final pixel colors, showing as noise in the image. In this paper, we describe a series of improvements to stochastic transparency that enable stratified sampling in both spatial and alpha domains. As a result, the amount of noise in the image is significantly reduced, while the result remains unbiased.

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.

In this paper, we examine the possibilities of using voxel representations as a generic way for expressing complex and feature-rich geometry on current and future GPUs. We present in detail a compact data structure for storing voxels and an efficient algorithm for performing ray casts using this structure. We augment the voxel data with novel contour information that increases geometric resolution, allows more compact encoding of smooth surfaces, and accelerates ray casts. We also employ a novel normal compression format for storing high-precision object-space normals. Finally, we present a variable-radius postprocess filtering technique for smoothing out blockiness caused by discrete sampling of shading attributes. Based on benchmark results, we show that our voxel representation is competitive with triangle-based representations in terms of ray casting performance, while allowing tremendously greater geometric detail and unique shading information for every voxel. Our voxel codebase is open sourced and available at http://code.google.com/p/efficient-sparse-voxel-octrees/.

In this paper we examine the possibilities of using voxel representations as a generic way for expressing complex and feature-rich geometry on current and future GPUs. We present in detail a compact data structure for storing voxels and an efficient algorithm for performing ray casts using this structure.

Abstract Ambient occlusion has proven to be a useful tool for producing realistic images, both in offline rendering and interactive applications. In production rendering, ambient occlusion is typically computed by casting a large number of short shadow rays from each visible point, yielding unparalleled quality but long rendering times. Interactive applications typically use screen‐space approximations which are fast but suffer from systematic errors due to missing information behind the nearest depth layer. In this paper, we present two efficient methods for calculating ambient occlusion so that the results match those produced by a ray tracer. The first method is targeted for rasterization‐based engines, and it leverages the GPU graphics pipeline for finding occlusion relations between scene triangles and the visible points. The second method is a drop‐in replacement for ambient occlusion computation in offline renderers, allowing the querying of ambient occlusion for any point in the scene. Both methods are based on the principle of simultaneously computing the result of all shadow rays for a single receiver point.