Video editing with diffusion models presents significant challenges, especially under real-time constraints. Current methods either enhance temporal consistency at the cost of slow processing or rely on frame-by-frame editing, leading to flickering and temporal artifacts. To address both challenges, we propose SmoothDiffusion-VE, a streaming-based editing approach that improves temporal consistency and processing speed through our proposed Adaptive Feature Cache (AFC) and motion-guided attention. The AFC dynamically adjusts the caching behavior based on perceptual similarity (LPIPS) between frames, i.e., shifting to a mini-cache mode for similar frames to reduce computational load. Conversely, significant frame changes trigger deeper caching to maintain robust temporal coherence. Our motion-guided attention selectively focuses on dynamic regions using optical flow, reducing unnecessary computations in static areas and accelerating processing. SmoothDiffusion-VE can run 28 FPS on one RTX 4090 GPU, achieving a 1564× speedup over Plug-and-Play Diffusion (PNP) and a 1916× speedup over Diffusion Motion Transfer (DMT), delivering a powerful solution for fast and consistent video editing.

Extended reality (XR) devices, including augmented, virtual, and mixed reality, provide a deeply immersive experience. However, practical limitations like weight, heat, and comfort put extreme constraints on the performance, power consumption, and image quality of such systems. In this paper, we study how these constraints form the tradeoff between Fixed Foveated Rendering (FFR), Gaze-Tracked Foveated Rendering (TFR), and conventional, non-foveated rendering. While existing papers have often studied these methods, we provide the first comprehensive study of their relative feasibility in practical systems with limited battery life and computational budget. We show that TFR with the added cost of the gaze-tracker can often be more expensive than FFR. Thus, we co-design a gaze-tracked foveated renderer considering its benefits in computation, power efficiency, and tradeoffs in image quality. We describe principled approximations for eye tracking which provide up to a 9x speedup in runtime performance with approximately a 20x improvement in energy efficiency when run on a mobile GPU. In isolation, these approximations appear to significantly degrade the gaze quality, but appropriate compensation in the visual pipeline can mitigate the loss. Overall, we show that with a highly optimized gaze-tracker, TFR is feasible compared to FFR, resulting in up to 1.25x faster frame times while also reducing total energy consumption by over 40%.

Speed and consistency of target-shifting play a crucial role in human ability to perform complex tasks. Shifting our gaze between objects of interest quickly and consistently requires changes both in depth and direction. Gaze changes in depth are driven by slow, inconsistent vergence movements which rotate the eyes in opposite directions, while changes in direction are driven by ballistic, consistent movements called saccades , which rotate the eyes in the same direction. In the natural world, most of our eye movements are a combination of both types. While scientific consensus on the nature of saccades exists, vergence and combined movements remain less understood and agreed upon. We eschew the lack of scientific consensus in favor of proposing an operationalized computational model which predicts the completion time of any type of gaze movement during target-shifting in 3D. To this end, we conduct a psychophysical study in a stereo VR environment to collect more than 12,000 gaze movement trials, analyze the temporal distribution of the observed gaze movements, and fit a probabilistic model to the data. We perform a series of objective measurements and user studies to validate the model. The results demonstrate its predictive accuracy, generalization, as well as applications for optimizing visual performance by altering content placement. Lastly, we leverage the model to measure differences in human target-changing time relative to the natural world, as well as suggest scene-aware projection depth. By incorporating the complexities and randomness of human oculomotor control, we hope this research will support new behavior-aware metrics for VR/AR display design, interface layout, and gaze-contingent rendering.

Previous blind or No Reference (NR) Image / video quality assessment (IQA/VQA) models largely rely on features drawn from natural scene statistics (NSS), but under the assumption that the image statistics are stationary in the spatial domain. Several of these models are quite successful on standard pictures. However, in Virtual Reality (VR) applications, foveated video compression is regaining attention, and the concept of space-variant quality assessment is of interest, given the availability of increasingly high spatial and temporal resolution contents and practical ways of measuring gaze direction. Distortions from foveated video compression increase with increased eccentricity, implying that the natural scene statistics are space-variant. Towards advancing the development of foveated compression / streaming algorithms, we have devised a no-reference (NR) foveated video quality assessment model, called FOVQA, which is based on new models of space-variant natural scene statistics (NSS) and natural video statistics (NVS). Specifically, we deploy a space-variant generalized Gaussian distribution (SV-GGD) model and a space-variant asynchronous generalized Gaussian distribution (SV-AGGD) model of mean subtracted contrast normalized (MSCN) coefficients and products of neighboring MSCN coefficients, respectively. We devise a foveated video quality predictor that extracts radial basis features, and other features that capture perceptually annoying rapid quality fall-offs. We find that FOVQA achieves state-of-the-art (SOTA) performance on the new 2D LIVE-FBT-FCVR database, as compared with other leading Foveated IQA / VQA models. we have made our implementation of FOVQA available at: https://live.ece.utexas.edu/research/Quality/FOVQA.zip.

We aim to ask and answer an essential question " how quickly do we react after observing a displayed visual target?" To this end, we present psychophysical studies that characterize the remarkable disconnect between human saccadic behaviors and spatial visual acuity. Building on the results of our studies, we develop a perceptual model to predict temporal gaze behavior, particularly saccadic latency, as a function of the statistics of a displayed image. Specifically, we implement a neurologically-inspired probabilistic model that mimics the accumulation of confidence that leads to a perceptual decision. We validate our model with a series of objective measurements and user studies using an eye-tracked VR display. The results demonstrate that our model prediction is in statistical alignment with real-world human behavior. Further, we establish that many sub-threshold image modifications commonly introduced in graphics pipelines may significantly alter human reaction timing, even if the differences are visually undetectable. Finally, we show that our model can serve as a metric to predict and alter reaction latency of users in interactive computer graphics applications, thus may improve gaze-contingent rendering, design of virtual experiences, and player performance in e-sports. We illustrate this with two examples: estimating competition fairness in a video game with two different team colors, and tuning display viewing distance to minimize player reaction time.

In Virtual Reality (VR) systems, head mounted displays (HMDs) are widely used to present VR contents. When displaying immersive (360° video) scenes, greater challenges arise due to limitations of computing power, frame rate, and transmission bandwidth. To address these problems, a variety of foveated video compression and streaming methods have been proposed, which seek to exploit the nonuniform sampling density of the retinal photoreceptors and ganglion cells, which decreases rapidly with increasing eccentricity. Creating foveated immersive video content leads to the need for specialized foveated video quality pridictors. Here we propose a No-Reference (NR or blind) method which we call “Space-Variant BRISQUE (SV-BRISQUE),” which is based on a new space-variant natural scene statistics model. When tested on a large database of foveated, compression-distorted videos along with human opinions of them, we found that our new model algorithm achieves state of the art (SOTA) performance with correlation 0.88 / 0.90 (PLCC / SROCC) against human subjectivity.

Traditional video codecs follow the predictive coding architecture of motion-compensated prediction and residual transform coding. Inspired by recent advances in deep learning, we propose a new deep learning video compression architecture that does not require motion estimation, which is the most expensive component in traditional video codecs. Our network consists of three components: a Displacement Calculation Unit (DCU), a Displacement Compression Network (DCN), and a Frame Reconstruction Network (FRN). The DCU exploits displaced frame differences as motion information, thus removing the need for motion estimation found in hybrid codecs. DCN utilizes an RNN-based network to learn temporal dependencies between frames. In the FRN, a new version of the UNet model, called LSTM-UNet is proposed and utilized to learn space-time differential representations of the videos. Our experimental results show that our compression model, MOtionless VIdeo Codec (MOVI-Codec), learns how to efficiently compress videos without computing motion and outperforms the video coding standard H.264 and exceeds the performance of the modern global standard HEVC codec as measured by MS-SSIM, especially on higher resolution videos.

Virtual Reality is regaining attention due to recent advancements in hardware technology. Immersive images / videos are becoming widely adopted to carry omnidirectional visual information. However, due to the requirements for higher spatial and temporal resolution of real video data, immersive videos require significantly larger bandwidth consumption. To reduce stresses on bandwidth, foveated video compression is regaining popularity, whereby the space-variant spatial resolution of the retina is exploited. Towards advancing the progress of foveated video compression, we propose a full reference (FR) foveated image quality assessment algorithm, which we call foveated entropic differencing (FED), which employs the natural scene statistics of bandpass responses by applying differences of local entropies weighted by a foveation-based error sensitivity function. We evaluate the proposed algorithm by measuring the correlations of the predictions that FED makes against human judgements on the newly created 2D and 3D LIVE-FBT-FCVR databases for Virtual Reality (VR). The performance of the proposed algorithm yields state-of-the-art as compared with other existing full reference algorithms. Software for FED has been made available at: http://live.ece.utexas.edu/research/Quality/FED.zip

In Virtual Reality (VR), the requirements of much higher resolution and smooth viewing experiences under rapid and often real-time changes in viewing direction, leads to significant challenges in compression and communication. To reduce the stresses of very high bandwidth consumption, the concept of foveated video compression is being accorded renewed interest. By exploiting the space-variant property of retinal visual acuity, foveation has the potential to substantially reduce video resolution in the visual periphery, with hardly noticeable perceptual quality degradations. Accordingly, foveated image / video quality predictors are also becoming increasingly important, as a practical way to monitor and control future foveated compression algorithms. Towards advancing the development of foveated image / video quality assessment (FIQA / FVQA) algorithms, we have constructed 2D and (stereoscopic) 3D VR databases of foveated / compressed videos, and conducted a human study of perceptual quality on each database. Each database includes 10 reference videos and 180 foveated videos, which were processed by 3 levels of foveation on the reference videos. Foveation was applied by increasing compression with increased eccentricity. In the 2D study, each video was of resolution 7680×3840 and was viewed and quality-rated by 36 subjects, while in the 3D study, each video was of resolution 5376×5376 and rated by 34 subjects. Both studies were conducted on top of a foveated video player having low motion-to-photon latency (~50ms). We evaluated different objective image and video quality assessment algorithms, including both FIQA / FVQA algorithms and non-foveated algorithms, on our so called LIVE-Facebook Technologies Foveation-Compressed Virtual Reality (LIVE-FBT-FCVR) databases. We also present a statistical evaluation of the relative performances of these algorithms. The LIVE-FBT-FCVR databases have been made publicly available and can be accessed at https://live.ece.utexas.edu/research/LIVEFBTFCVR/index.html.

FovVideoVDP is a video difference metric that models the spatial, temporal, and peripheral aspects of perception. While many other metrics are available, our work provides the first practical treatment of these three central aspects of vision simultaneously. The complex interplay between spatial and temporal sensitivity across retinal locations is especially important for displays that cover a large field-of-view, such as Virtual and Augmented Reality displays, and associated methods, such as foveated rendering. Our metric is derived from psychophysical studies of the early visual system, which model spatio-temporal contrast sensitivity, cortical magnification and contrast masking. It accounts for physical specification of the display (luminance, size, resolution) and viewing distance. To validate the metric, we collected a novel foveated rendering dataset which captures quality degradation due to sampling and reconstruction. To demonstrate our algorithm's generality, we test it on 3 independent foveated video datasets, and on a large image quality dataset, achieving the best performance across all datasets when compared to the state-of-the-art.

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.

Abstract Despite recent advances in Monte Carlo path tracing at interactive rates, denoised image sequences generated with few samples per‐pixel often yield temporally unstable results and loss of high‐frequency details. We present a novel adaptive rendering method that increases temporal stability and image fidelity of low sample count path tracing by distributing samples via spatio‐temporal joint optimization of sampling and denoising. Adding temporal optimization to the sample predictor enables it to learn spatio‐temporal sampling strategies such as placing more samples in disoccluded regions, tracking specular highlights, etc; adding temporal feedback to the denoiser boosts the effective input sample count and increases temporal stability. The temporal approach also allows us to remove the initial uniform sampling step typically present in adaptive sampling algorithms. The sample predictor and denoiser are deep neural networks that we co‐train end‐to‐end over multiple consecutive frames. Our approach is scalable, allowing trade‐off between quality and performance, and runs at near real‐time rates while achieving significantly better image quality and temporal stability than previous methods.

Blue noise sampling has proved useful for many graphics applications, but remains underexplored in high-dimensional spaces due to the difficulty of generating distributions and proving properties about them. We present a blue noise sampling method with good quality and performance across different dimensions. The method, spoke-dart sampling, shoots rays from prior samples and selects samples from these rays. It combines the advantages of two major high-dimensional sampling methods: the locality of advancing front with the dimensionality-reduction of hyperplanes, specifically line sampling. We prove that the output sampling is saturated with high probability, with bounds on distances between pairs of samples and between any domain point and its nearest sample. We demonstrate spoke-dart applications for approximate Delaunay graph construction, global optimization, and robotic motion planning. Both the blue-noise quality of the output distribution and the adaptability of the intermediate processes of our method are useful in these applications.

Redirected walking techniques can enhance the immersion and visual-vestibular comfort of virtual reality (VR) navigation, but are often limited by the size, shape, and content of the physical environments. We propose a redirected walking technique that can apply to small physical environments with static or dynamic obstacles. Via a head- and eye-tracking VR headset, our method detects saccadic suppression and redirects the users during the resulting temporary blindness. Our dynamic path planning runs in real-time on a GPU, and thus can avoid static and dynamic obstacles, including walls, furniture, and other VR users sharing the same physical space. To further enhance saccadic redirection, we propose subtle gaze direction methods tailored for VR perception. We demonstrate that saccades can significantly increase the rotation gains during redirection without introducing visual distortions or simulator sickness. This allows our method to apply to large open virtual spaces and small physical environments for room-scale VR. We evaluate our system via numerical simulations and real user studies.

Foveated rendering is a performance optimization based on the well-known degradation of peripheral visual acuity. It reduces computational costs by showing a high-quality image in the user’s central (foveal) vision and a lower quality image in the periphery. Foveated rendering is a promising optimization for Virtual Reality (VR) graphics, and generally requires accurate and low-latency eye tracking to ensure correctness even when a user makes large, fast eye movements such as saccades. However, due to the phenomenon of saccadic omission, it is possible that these requirements may be relaxed. In this article, we explore the effect of latency for foveated rendering in VR applications. We evaluated the detectability of visual artifacts for three techniques capable of generating foveated images and for three different radii of the high-quality foveal region. Our results show that larger foveal regions allow for more aggressive foveation, but this effect is more pronounced for temporally stable foveation techniques. Added eye tracking latency of 80--150ms causes a significant reduction in acceptable amount of foveation, but a similar decrease in acceptable foveation was not found for shorter eye-tracking latencies of 20--40ms, suggesting that a total system latency of 50--70ms could be tolerated.

Foveated rendering synthesizes images with progressively less detail outside the eye fixation region, potentially unlocking significant speedups for wide field-of-view displays, such as head mounted displays, where target framerate and resolution is increasing faster than the performance of traditional real-time renderers. To study and improve potential gains, we designed a foveated rendering user study to evaluate the perceptual abilities of human peripheral vision when viewing today's displays. We determined that filtering peripheral regions reduces contrast, inducing a sense of tunnel vision. When applying a postprocess contrast enhancement, subjects tolerated up to 2× larger blur radius before detecting differences from a non-foveated ground truth. After verifying these insights on both desktop and head mounted displays augmented with high-speed gaze-tracking, we designed a perceptual target image to strive for when engineering a production foveated renderer. Given our perceptual target, we designed a practical foveated rendering system that reduces number of shades by up to 70% and allows coarsened shading up to 30° closer to the fovea than Guenter et al. [2012] without introducing perceivable aliasing or blur. We filter both pre- and post-shading to address aliasing from undersampling in the periphery, introduce a novel multiresolution- and saccade-aware temporal antialising algorithm, and use contrast enhancement to help recover peripheral details that are resolvable by our eye but degraded by filtering. We validate our system by performing another user study. Frequency analysis shows our system closely matches our perceptual target. Measurements of temporal stability show we obtain quality similar to temporally filtered non-foveated renderings.

We present Piko, a framework for designing, optimizing, and retargeting implementations of graphics pipelines on multiple architectures. Piko programmers express a graphics pipeline by organizing the computation within each stage into spatial bins and specifying a scheduling preference for these bins. Our compiler, Pikoc, compiles this input into an optimized implementation targeted to a massively-parallel GPU or a multicore CPU. Piko manages work granularity in a programmable and flexible manner, allowing programmers to build load-balanced parallel pipeline implementations, to exploit spatial and producer-consumer locality in a pipeline implementation, and to explore tradeoffs between these considerations. We demonstrate that Piko can implement a wide range of pipelines, including rasterization, Reyes, ray tracing, rasterization/ray tracing hybrid, and deferred rendering. Piko allows us to implement efficient graphics pipelines with relative ease and to quickly explore design alternatives by modifying the spatial binning configurations and scheduling preferences for individual stages, all while delivering real-time performance that is within a factor six of state-of-the-art rendering systems.

We formalize sampling a function using k -d darts. A k -d Dart is a set of independent, mutually orthogonal, k -dimensional hyperplanes called k -d flats. A dart has d choose k flats, aligned with the coordinate axes for efficiency. We show k -d darts are useful for exploring a function's properties, such as estimating its integral, or finding an exemplar above a threshold. We describe a recipe for converting some algorithms from point sampling to k -d dart sampling, if the function can be evaluated along a k -d flat. We demonstrate that k -d darts are more efficient than point-wise samples in high dimensions, depending on the characteristics of the domain: for example, the subregion of interest has small volume and evaluating the function along a flat is not too expensive. We present three concrete applications using line darts (1-d darts): relaxed maximal Poisson-disk sampling, high-quality rasterization of depth-of-field blur, and estimation of the probability of failure from a response surface for uncertainty quantification. Line darts achieve the same output fidelity as point sampling in less time. For Poisson-disk sampling, we use less memory, enabling the generation of larger point distributions in higher dimensions. Higher-dimensional darts provide greater accuracy for a particular volume estimation problem.

Abstract We provide a simple algorithm and data structures for d‐dimensional unbiased maximal Poisson‐disk sampling. We use an order of magnitude less memory and time than the alternatives. Our results become more favorable as the dimension increases. This allows us to produce bigger samplings. Domains may be non‐convex with holes. The generated point cloud is maximal up to round‐off error. The serial algorithm is provably bias‐free. For an output sampling of size n in fixed dimension d, we use a linear memory budget and empirical θ(n) runtime. No known methods scale well with dimension, due to the “curse of dimensionality.” The serial algorithm is practical in dimensions up to 5, and has been demonstrated in 6d. We have efficient GPU implementations in 2d and 3d. The algorithm proceeds through a finite sequence of uniform grids. The grids guide the dart throwing and track the remaining disk‐free area. The top‐level grid provides an efficient way to test if a candidate dart is disk‐free. Our uniform grids are like quadtrees, except we delay splits and refine all leaves at once. Since the quadtree is flat it can be represented using very little memory: we just need the indices of the active leaves and a global level. Also it is very simple to sample from leaves with uniform probability.

In this paper we explore strategies for screen-space tiling in programmable rendering pipelines on modern GPUs. The need for tiling arises from the inherent unpredictability of workload size in many rendering pipelines. While it additionally also reduces the bandwidth demand of the rendering operations and increases locality of memory accesses, we focus on meeting resource limits in programmable rendering. We show that by ensuring load balance, dynamic tiling in a modern programmable GPU pipeline delivers performance improvements over traditional static tiling techniques. We explore prediction and splitting algorithms for adaptive tiling schemes and demonstrate that even approximate estimates of tile bounds can reduce tiling overhead. We evaluate the above optimizations by implementing them for a three GPU-based tiled renderers while keeping our observations general enough so that they are applicable to other forms of rendering as well.

We solve the problem of generating a uniform Poisson-disk sampling that is both maximal and unbiased over bounded non-convex domains. To our knowledge this is the first provably correct algorithm with time and space dependent only on the number of points produced. Our method has two phases, both based on classical dart-throwing. The first phase uses a background grid of square cells to rapidly create an unbiased, near-maximal covering of the domain. The second phase completes the maximal covering by calculating the connected components of the remaining uncovered voids, and by using their geometry to efficiently place unbiased samples that cover them. The second phase converges quickly, overcoming a common difficulty in dart-throwing methods. The deterministic memory is O ( n ) and the expected running time is O ( n log n ), where n is the output size, the number of points in the final sample. Our serial implementation verifies that the log n dependence is minor, and nearly O ( n ) performance for both time and memory is achieved in practice. We also present a parallel implementation on GPUs to demonstrate the parallel-friendly nature of our method, which achieves 2.4x the performance of our serial version.

Abstract We present a strategy for parallelizing the composite and filter operations suitable for an order‐independent rendering pipeline implemented on a modern graphics processor. Conventionally, this task is parallelized across pixels/subpixels, but serialized along individual depth layers. However, our technique extends the domain of parallelization to individual fragments (samples), avoiding a serial dependence on the number of depth layers, which can be a constraint for scenes with high depth complexity. As a result, our technique scales with the number of fragments and can sustain a consistent and predictable throughput in scenes with both low and high depth complexity, including those with a high variability of depth complexity within a single frame. We demonstrate composite/filter performance in excess of 50M fragments/sec for scenes with more than 1500 semi‐transparent layers.

With increasing interest in sophisticated graphics capabilities in mobile systems, energy consumption of graphics hardware is becoming a major design concern in addition to the traditional performance enhancement criteria. Among the different steps in the graphics processing pipeline, we have observed that memory accesses during texture mapping - a highly memory intensive phase - contribute 30–40% of the energy consumed in typical embedded graphics processors. This makes the texture mapping subsystem an attractive candidate for energy optimization. We argue that a standard cache hierarchy, commonly used by researchers and commercial graphics processors for texture mapping, is wasteful of energy, and propose the Texture Filter Memory, an energy efficient architecture that exploits locality and the relatively high degree of predictability in texture memory access patterns. Our architecture consumes 75% lesser energy for texturing in a fixed function pipeline, incurring no performance overhead and a small area overhead over conventional texture mapping hardware.

We present a technique for designing memory-bound algorithms with high data reuse on Graphics Processing Units (GPUs) equipped with close-to-ALU software-managed memory. The approach is based on the efficient use of this memory through the implementation of a software-managed cache. We also present an analytical model for performance analysis of such algorithms.

We present a GPU based implementation of Reyes-style adaptive surface subdivision, known in Reyes terminology as the Bound/Split and Dice stages. The performance of this task is important for the Reyes pipeline to map efficiently to graphics hardware, but its recursive nature and irregular and unbounded memory requirements present a challenge to an efficient implementation. Our solution begins by characterizing Reyes subdivision as a work queue with irregular computation, targeted to a massively parallel GPU. We propose efficient solutions to these general problems by casting our solution in terms of the fundamental primitives of prefix-sum and reduction, often encountered in parallel and GPGPU environments. Our results indicate that real-time Reyes subdivision can indeed be obtained on today's GPUs. We are able to subdivide a complex model to subpixel accuracy within 15 ms. Our measured performance is several times better than that of Pixar's RenderMan. Our implementation scales well with the input size and depth of subdivision. We also address concerns of memory size and bandwidth, and analyze the feasibility of conventional ideas on screen-space buckets.