Abstract: A modular architecture is provided for denoising Monte Carlo renderings using neural networks. The temporal approach extracts and combines feature representations from neighboring frames rather than building a temporal context using recurrent connections. A multiscale architecture includes separate single-frame or temporal denoising modules for individual scales, and one or more scale compositor neural networks configured to adaptively blend individual scales. An error-predicting module is configured to produce adaptive sampling maps for a renderer to achieve more uniform residual noise distribution. An asymmetric loss function may be used for training the neural networks, which can provide control over the variance-bias trade-off during denoising.

Abstract: Embodiments provide for three-dimensional cloth modeling are provided. A first mesh comprising a plurality of faces defined by a plurality of edges is accessed, and a render mesh is generated using quadrangulated tessellation of the first mesh, where the render mesh comprises quad faces. One or more attributes of the plurality of faces of the first mesh are transferred to one or more of the quad faces of the render mesh using a stochastic transfer operation. The render mesh is displayed via a graphical user interface (GUI).

Abstract: Techniques for aligning object representations for animation include analyzing a source object representation and a target object representation to identify a category of the source object and a category of the target object. Based on the category or categories of the objects, a feature extraction machine learning models is selected. The source object representation and the target object representation are provided as input to the selected feature extraction machine learning model to generate respective semantic descriptors and shape vectors for the source and target objects. Based on the semantic descriptors and the shape vectors for the source and target objects, an alignment machine learning model generates an aligned target object representation that is aligned with the source object representation and usable for animating the target object.

Abstract: A method performed by a computer is disclosed. The method comprises receiving color data for input pixels of an input image and an input set of features used to render the input image of a three-dimensional animation environment, wherein the input pixels are of a first resolution. The computer may then load into memory a generator of a generative adversarial network including a neural network used to scale the input image, the neural network trained using training data comprising color data of training input images and training output images and a training set of the features used to render the training input images. After the generator is loaded into memory, the computer may generate an output image having a second resolution that is different than the first resolution by passing the color data and the input set of features through the generator.

Abstract: Systems and methods generate a modified three-dimensional mesh representation of an object using a trained neural network. A computer system receives a set of input values for posing an initial mesh defining a surface of a three-dimensional object. The computer system provides the input values to a neural network trained on posed meshes generated using a rigging model to generate mesh offset values based upon the set of input values and the initial mesh. The neural network includes an input layer, an output layer, and a plurality of intermediate layers. The computer system generates, by the output layer of the neural network, a set of offset values corresponding to a set of three-dimensional target points based on the set of input values. The offset values are applied to the initial mesh to generate a posed mesh. The computer system outputs the posed mesh for generating an animation frame.

Abstract: Systems and methods automatically generate contours on an illustrated object for performing an animation. Contour lines are generated on the surface of the object according to criteria related to the shape of the surface of the object. Points of the contour lines that are occluded from a virtual camera are identified. The occluded points are removed to generate visible lines. The visible lines are extruded to define a three-dimensional volume defining contours of the object. The object itself, along with the three-dimensional volume, are illuminated and rendered. The parameters defining the opacity and color of the contour may differ from corresponding parameters of the rest of the object, so that the contours stand out and define portions of the object. The contours are useful in contexts such as defining areas of an object that is fuzzy or cloudy in appearance, as well as creating certain artistic effects.

Abstract: Systems and methods automatically generate contours on an illustrated object for performing an animation. Contour lines are generated on the surface of the object according to criteria related to the shape of the surface of the object. Points of the contour lines that are occluded from a virtual camera are identified. The occluded points are removed to generate visible lines. The visible lines are extruded to define a three-dimensional volume defining contours of the object. The object itself, along with the three-dimensional volume, are illuminated and rendered. The parameters defining the opacity and color of the contour may differ from corresponding parameters of the rest of the object, so that the contours stand out and define portions of the object. The contours are useful in contexts such as defining areas of an object that is fuzzy or cloudy in appearance, as well as creating certain artistic effects.

Abstract: A modular architecture is provided for denoising Monte Carlo renderings using neural networks. The temporal approach extracts and combines feature representations from neighboring frames rather than building a temporal context using recurrent connections. A multiscale architecture includes separate single-frame or temporal denoising modules for individual scales, and one or more scale compositor neural networks configured to adaptively blend individual scales. An error-predicting module is configured to produce adaptive sampling maps for a renderer to achieve more uniform residual noise distribution. An asymmetric loss function may be used for training the neural networks, which can provide control over the variance-bias trade-off during denoising.

Abstract: Provided are methods, systems, and computer-program products for recovering from intersections during a simulation of an animated scene when a collision detection operation is active. For example, the collision detection operation can be selectively activated and deactivated during the simulation of one or more objects for a time step based on an intersection analysis, which can identify intersections of the one or more objects for the time step. Once the collision detection operation is deactivated, a collision response can apply one or more forces to intersecting portions of the one or more objects to eliminate the intersections of the one or more objects. For example, a portion of a cloth that is in a state of intersection can be configured such that the collision detection operation is not performed on the portion, thereby allowing the cloth to be removed from inside of another object by a collision response algorithm.

Abstract: A modular architecture is provided for denoising Monte Carlo renderings using neural networks. The temporal approach extracts and combines feature representations from neighboring frames rather than building a temporal context using recurrent connections. A multiscale architecture includes separate single-frame or temporal denoising modules for individual scales, and one or more scale compositor neural networks configured to adaptively blend individual scales. An error-predicting module is configured to produce adaptive sampling maps for a renderer to achieve more uniform residual noise distribution. An asymmetric loss function may be used for training the neural networks, which can provide control over the variance-bias trade-off during denoising.

Abstract: Supervised machine learning using neural networks is applied to denoising images rendered by MC path tracing. Specialization of neural networks may be achieved by using a modular design that allows reusing trained components in different networks and facilitates easy debugging and incremental building of complex structures. Specialization may also be achieved by using progressive neural networks. In some embodiments, training of a neural-network based denoiser may use importance sampling, where more challenging patches or patches including areas of particular interests within a training dataset are selected with higher probabilities than others. In some other embodiments, generative adversarial networks (GANs) may be used for training a machine-learning based denoiser as an alternative to using pre-defined loss functions.

Abstract: Embodiments provide for cut-aware UV transfer. Embodiments include receiving a surface correspondence map that maps points of a source mesh to points of a target mesh. Embodiments include generating a set of functions encoding locations of seam curves and wrap curves from a source UV map of the source mesh. Embodiments include using the set of functions and the surface correspondence map to determine a target UV map that maps a plurality of target seam curves and a plurality of target wrap curves to the target mesh. Embodiments include transferring a two-dimensional parametrization of the source UV map to the target UV map.

Abstract: Embodiments provide for transferring mesh connectivity. Embodiments include receiving a definition of a correspondence between a first curve for a source mesh and a second curve for a target shape. Embodiments include initializing an output mesh by setting a third plurality of vertices in the output mesh equal to a first plurality of vertices in the source mesh. Embodiments include transforming the output mesh by modifying the third plurality of vertices based on the first curve, the second curve, and a second plurality of vertices of the target mesh. Vertices of the third plurality of vertices that relate to the first curve are conformed to a shape defined by the second curve, and vertex modifications that result in affine transformations of faces in the output mesh are favored. Embodiments include using the output mesh to transfer an attribute from the source mesh to the target shape.

Abstract: The present disclosure relates to using a neural network to efficiently denoise images that were generated by a ray tracer. The neural network can be trained using noisy images generated with noisy samples and corresponding denoised or high-sampled images (e.g., many random samples). An input feature to the neural network can include color from pixels of an image. Other input features to the neural network, which would not be known in normal image processing, can include shading normal, depth, albedo, and other characteristics available from a computer-generated scene. After the neural network is trained, a noisy image that the neural network has not seen before can have noise removed without needing manual intervention.

Abstract: Techniques are disclosed that allow animators to easily share and reuse character poses such as gestures, expressions, and mouth shapes. When starting on a new shot, an animator often wants a character to have the same pose exactly as the end of the previous shot. According to various embodiments, an animator can easily set up these hookup poses by animator copying a pose directly from a clip of prerecorded media. In one aspect, a pose at the current playhead of the playback tool is copied into a software buffer of an animation tool and then pasted into a character. Thus, the animator may copy a pose exactly as he/she is seeing visually. In various aspects, animators can choose a pose from an entire inventory of available animated videos. This provides a more efficient method for selecting a pose since the user can easily choose and pick a pose from a large inventory of animated videos and bring in a desired pose in a matter of a few mouse clicks.

Abstract: Embodiments provide for sculpt transfer. Embodiments include identifying a source polygon of a source mesh that corresponds to a target polygon of a target mesh. Embodiments include determining a first matrix defining a first rotation that aligns a target rest state of the target polygon to a source rest state of the source polygon, determining a second matrix defining a linear transformation that aligns the source rest state to a source pose of the source polygon, wherein the linear transformation comprises rotating and stretching, determining a third matrix defining a second rotation that aligns the source pose to the target rest state, and determining a fourth matrix defining a third rotation that aligns the source rest state to the source pose. Embodiments include determining a target pose of the target polygon based on the target rest state, the first matrix, the second matrix, the third matrix, and the fourth matrix.

Abstract: Supervised machine learning using neural networks is applied to denoising images rendered by MC path tracing. Specialization of neural networks may be achieved by using a modular design that allows reusing trained components in different networks and facilitates easy debugging and incremental building of complex structures. Specialization may also be achieved by using progressive neural networks. In some embodiments, training of a neural-network based denoiser may use importance sampling, where more challenging patches or patches including areas of particular interests within a training dataset are selected with higher probabilities than others. In some other embodiments, generative adversarial networks (GANs) may be used for training a machine-learning based denoiser as an alternative to using pre-defined loss functions.

Abstract: Surface relaxation techniques are disclosed for smoothing the shapes of three-dimensional (3D) virtual geometry. In one embodiment, a surface relaxation application determines, for each of a number of vertices of a 3D virtual geometry, span-aware weights for each edge incident to the vertex based on the alignment of other edges incident to the vertex with an orthonormal frame of the edge constructed using a decal map. The surface relaxation application uses such span-aware weights to compute weighted averages that provide surface relaxation offsets. Further, the surface relaxation application may restore relaxation offsets from an original to a deformed geometry by determining relaxation offsets for both geometries and transferring the relaxation offsets from the original to the deformed 3D geometry using a blending of the determined relaxation offsets and a rotation. In another embodiment, volume is preserved by computing relaxation offsets in the plane and lifting relaxed vertices back to 3D.

Abstract: A modular architecture is provided for denoising Monte Carlo renderings using neural networks. The temporal approach extracts and combines feature representations from neighboring frames rather than building a temporal context using recurrent connections. A multiscale architecture includes separate single-frame or temporal denoising modules for individual scales, and one or more scale compositor neural networks configured to adaptively blend individual scales. An error-predicting module is configured to produce adaptive sampling maps for a renderer to achieve more uniform residual noise distribution. An asymmetric loss function may be used for training the neural networks, which can provide control over the variance-bias trade-off during denoising.

Abstract: A modular architecture is provided for denoising Monte Carlo renderings using neural networks. The temporal approach extracts and combines feature representations from neighboring frames rather than building a temporal context using recurrent connections. A multiscale architecture includes separate single-frame or temporal denoising modules for individual scales, and one or more scale compositor neural networks configured to adaptively blend individual scales. An error-predicting module is configured to produce adaptive sampling maps for a renderer to achieve more uniform residual noise distribution. An asymmetric loss function may be used for training the neural networks, which can provide control over the variance-bias trade-off during denoising.