Abstract: A neural network-based rendering technique increases temporal stability and image fidelity of low sample count path tracing by optimizing a distribution of samples for rendering each image in a sequence. A sample predictor neural network learns spatio-temporal sampling strategies such as placing more samples in dis-occluded regions and tracking specular highlights. Temporal feedback enables a denoiser neural network to boost the effective input sample count and increases temporal stability. The initial uniform sampling step typically present in adaptive sampling algorithms is not needed. The sample predictor and denoiser operate at interactive rates to achieve significantly improved image quality and temporal stability compared with conventional adaptive sampling techniques.

Abstract: A method, computer readable medium, and system are disclosed for temporally stable data reconstruction. A sequence of input data including artifacts is received. A first input data frame is processed using layers of a neural network model to produce external state including a reconstructed first data frame that approximates the first input data frame without artifacts. Hidden state generated during processing of the first input data is not provided as an input to the layer to process second input data. The external state is warped, using difference data corresponding to changes between input data frames, to produce warped external state more closely aligned with the second input data frame. The second input data frame is processed, based on the warped external state, using the layers of the neural network model to produce a reconstructed second data frame that approximates the second data frame without artifacts.

Abstract: Motivated by the ability of humans to quickly and accurately detect visual artifacts in images, a neural network model is trained to identify and locate visual artifacts in a sequence of rendered images without comparing the sequence of rendered images against a ground truth reference. Examples of visual artifacts include aliasing, blurriness, mosaicking, and overexposure. The neural network model provides a useful fully-automated tool for evaluating the quality of images produced by rendering systems. The neural network model may be trained to evaluate the quality of images for video processing, encoding, and/or compression techniques. In an embodiment, the sequence includes at least four images corresponding to a video or animation.

Abstract: The technology disclosed herein involves using a machine learning model (e.g., CNN) to expand lower dynamic-range image content (e.g., SDR images) into higher dynamic-range image content (e.g., HDR images). The machine learning model can take as input the lower dynamic-range image and can output multiple expansion maps that are used to make the expanded image appear more natural. The expansion maps may be used by image operators to smooth color banding and to dim overexposed regions or user interface elements in the expanded image. The expanded content (e.g., HDR image content) may then be provided to one or more devices for display or storage.

Abstract: In one embodiment, a method includes accessing a pair of stereo images for a scene, where each image of the pair of stereo images has incomplete pixel information and k channels, stacking the pair of stereo images to form a stacked input image with 2k channels, processing the stacked input image using a machine-learning model to generate a stacked output image with 2k channels, and separating the stacked output image with 2k channels into a pair of reconstructed stereo images for the scene, where each image of the pair of reconstructed stereo images has complete pixel information and k channels.

Abstract: In one embodiment, a method includes determining characteristics of one or more areas in an image by analyzing pixels in the image, computing a sampling density for each of the one or more areas in the image based on the characteristics of the one or more areas, generating samples corresponding to the image by sampling pixels in each of the one or more areas according to the associated sampling density, and providing the samples to a machine-learning model as an input, where the machine-learning model is configured to reconstruct the image by processing the samples.

Abstract: In one embodiment, a computing system may access a video including a first frame and a second frame. The computing system may determine first sampling locations for the first frame and determine second sampling locations for the second frame by transforming the first sampling locations to the second frame according to an optical flow between the first frame and the second frame. The computing system may detect one or more invalid second sampling locations based on determining pixels in the first frame corresponding to the first sampling locations do not match pixels in the second frame corresponding to the second sampling locations. The computing system may reject the one or more invalid second sampling locations to determine third sampling locations for the second frame. The computing system may generate a sample of the video.

Abstract: A neural network-based rendering technique increases temporal stability and image fidelity of low sample count path tracing by optimizing a distribution of samples for rendering each image in a sequence. A sample predictor neural network learns spatio-temporal sampling strategies such as placing more samples in dis-occluded regions and tracking specular highlights. Temporal feedback enables a denoiser neural network to boost the effective input sample count and increases temporal stability. The initial uniform sampling step typically present in adaptive sampling algorithms is not needed. The sample predictor and denoiser operate at interactive rates to achieve significantly improved image quality and temporal stability compared with conventional adaptive sampling techniques.

Abstract: A neural network-based rendering technique increases temporal stability and image fidelity of low sample count path tracing by optimizing a distribution of samples for rendering each image in a sequence. A sample predictor neural network learns spatio-temporal sampling strategies such as placing more samples in dis-occluded regions and tracking specular highlights. Temporal feedback enables a denoiser neural network to boost the effective input sample count and increases temporal stability. The initial uniform sampling step typically present in adaptive sampling algorithms is not needed. The sample predictor and denoiser operate at interactive rates to achieve significantly improved image quality and temporal stability compared with conventional adaptive sampling techniques.

Abstract: A neural network-based rendering technique increases temporal stability and image fidelity of low sample count path tracing by optimizing a distribution of samples for rendering each image in a sequence. A sample predictor neural network learns spatio-temporal sampling strategies such as placing more samples in dis-occluded regions and tracking specular highlights. Temporal feedback enables a denoiser neural network to boost the effective input sample count and increases temporal stability. The initial uniform sampling step typically present in adaptive sampling algorithms is not needed. The sample predictor and denoiser operate at interactive rates to achieve significantly improved image quality and temporal stability compared with conventional adaptive sampling techniques.

Abstract: In one embodiment, a method includes projecting a source image onto a surface using a lens approximation component, where the surface is associated with sampling points approximating photoreceptors of an eye, where each sampling point has a corresponding photoreceptor type, sampling color information from the projected source image at the sampling points, where the color information sampled at each sampling point depends on the corresponding photoreceptor type, accessing pooling units approximating retinal ganglion cells (RGCs) of the eye, where each pooling unit is associated with groups of one or more of the sampling points, calculating weighted aggregations of the sampled color information associated with the groups of one or more sampling points associated with each pooling unit, and computing a perception profile for the source image based on the weighted aggregations associated with each of the pooling units.

Abstract: In one embodiment, a computing system may access a video including a first frame and a second frame. The computing system may determine first sampling locations for the first frame and determine second sampling locations for the second frame by transforming the first sampling locations to the second frame according to an optical flow between the first frame and the second frame. The computing system may select a subset of the second sampling locations based on a comparison between pixels in the first frame corresponding to the first sampling locations and pixels in the second frame corresponding to the second sampling locations. The computing system may define one or more rejection areas in the second frame based on the subset of the second sampling locations to determine third sampling locations in areas outside of the rejection areas. The computing system may generate a sample of the video.

Abstract: In one embodiment, a computing system may receive a video including a sequence of frames. The computing system may access a three-dimensional mask that specifies pixel-sampling locations, the three-dimensional mask having a first dimension and a second dimension corresponding to a spatial domain and a third dimension corresponding to a temporal domain. Blue noise property may be present in the pixel-sampling locations that are associated with each of a plurality of two-dimensional spatial slices of the three-dimensional mask in the spatial domain and the pixel-sampling locations that are associated with each of a plurality of one-dimensional temporal slices of the three-dimensional mask in the temporal domain. The computing system may generate a sample of the video by sampling the sequence of frames using the three-dimensional mask.

Abstract: In one embodiment, a method includes projecting a source image onto a surface using a lens approximation component, where the surface is associated with sampling points approximating photoreceptors of an eye, where each sampling point has a corresponding photoreceptor type, sampling color information from the projected source image at the sampling points, where the color information sampled at each sampling point depends on the corresponding photoreceptor type, accessing pooling units approximating retinal ganglion cells (RGCs) of the eye, where each pooling unit is associated with groups of one or more of the sampling points, calculating weighted aggregations of the sampled color information associated with the groups of one or more sampling points associated with each pooling unit, and computing a perception profile for the source image based on the weighted aggregations associated with each of the pooling units.

Abstract: In one embodiment, a computing system may access a video including a first frame and a second frame. The computing system may determine first sampling locations for the first frame and determine second sampling locations for the second frame by transforming the first sampling locations to the second frame according to an optical flow between the first frame and the second frame. The computing system may select a subset of the second sampling locations based on a comparison between pixels in the first frame corresponding to the first sampling locations and pixels in the second frame corresponding to the second sampling locations. The computing system may define one or more rejection areas in the second frame based on the subset of the second sampling locations to determine third sampling locations in areas outside of the rejection areas. The computing system may generate a sample of the video.

Abstract: In one embodiment, a computing system may receive a video including a sequence of frames. The computing system may access a three-dimensional mask that specifies pixel-sampling locations, the three-dimensional mask having a first dimension and a second dimension corresponding to a spatial domain and a third dimension corresponding to a temporal domain. Blue noise property may be present in the pixel-sampling locations that are associated with each of a plurality of two-dimensional spatial slices of the three-dimensional mask in the spatial domain and the pixel-sampling locations that are associated with each of a plurality of one-dimensional temporal slices of the three-dimensional mask in the temporal domain. The computing system may generate a sample of the video by sampling the sequence of frames using the three-dimensional mask.

Abstract: In one embodiment, a method includes accessing a pair of stereo images for a scene, where each image of the pair of stereo images has incomplete pixel information and k channels, stacking the pair of stereo images to form a stacked input image with 2k channels, processing the stacked input image using a machine-learning model to generate a stacked output image with 2k channels, and separating the stacked output image with 2k channels into a pair of reconstructed stereo images for the scene, where each image of the pair of reconstructed stereo images has complete pixel information and k channels.

Abstract: In one embodiment, a method includes determining characteristics of one or more areas in an image by analyzing pixels in the image, computing a sampling density for each of the one or more areas in the image based on the characteristics of the one or more areas, generating samples corresponding to the image by sampling pixels in each of the one or more areas according to the associated sampling density, and providing the samples to a machine-learning model as an input, where the machine-learning model is configured to reconstruct the image by processing the samples.

Abstract: Apparatuses, systems, and techniques used to train one or more neural networks to generate images comprising one or more features. In at least one embodiment, one or more neural networks are trained to determine one or more styles for an input image and then generate features associated with said one or more styles in an output image.

Abstract: The disclosed microtraining techniques improve accuracy of trained neural networks by performing iterative refinement at low learning rates using a relatively short series microtraining steps. A neural network training framework receives the trained neural network along with a second training dataset and set of hyperparameters. The neural network training framework produces a microtrained neural network by adjusting one or more weights of the trained neural network using a lower learning rate to facilitate incremental accuracy improvements without substantially altering the computational structure of the trained neural network. The microtrained neural network may be assessed for changes in accuracy and/or quality. Additional microtraining sessions may be performed on the microtrained neural network to further improve accuracy or quality.