Recently, there has been a surge of diverse methods for performing image editing by employing pre-trained unconditional generators. Applying these methods on real images, however, remains a challenge, as it necessarily requires the inversion of the images into their latent space. To successfully invert a real image, one needs to find a latent code that reconstructs the input image accurately, and more importantly, allows for its meaningful manipulation. In this paper, we carefully study the latent space of StyleGAN, the state-of-the-art unconditional generator. We identify and analyze the existence of a distortion-editability tradeoff and a distortion-perception tradeoff within the StyleGAN latent space. We then suggest two principles for designing encoders in a manner that allows one to control the proximity of the inversions to regions that StyleGAN was originally trained on. We present an encoder based on our two principles that is specifically designed for facilitating editing on real images by balancing these tradeoffs. By evaluating its performance qualitatively and quantitatively on numerous challenging domains, including cars and horses, we show that our inversion method, followed by common editing techniques, achieves superior real-image editing quality, with only a small reconstruction accuracy drop.
The task of age transformation illustrates the change of an individual's appearance over time. Accurately modeling this complex transformation over an input facial image is extremely challenging as it requires making convincing and possibly large changes to facial features and head shape, while still preserving the input identity. In this work, we present an image-to-image translation method that learns to directly encode real facial images into the latent space of a pre-trained unconditional GAN (e.g., StyleGAN) subject to a given aging shift. We employ a pre-trained age regression network used to explicitly guide the encoder in generating the latent codes corresponding to the desired age. In this formulation, our method approaches the continuous aging process as a regression task between the input age and desired target age, providing fine-grained control over the generated image. Moreover, unlike other approaches that operate solely in the latent space using a prior on the path controlling age, our method learns a more disentangled, non-linear path. Finally, we demonstrate that the end-to-end nature of our approach, coupled with the rich semantic latent space of StyleGAN, allows for further editing of the generated images. Qualitative and quantitative evaluations show the advantages of our method compared to state-of-the-art approaches.
State-of-the-art image-to-image translation methods tend to struggle in an imbalanced domain setting, where one image domain lacks richness and diversity. We introduce a new unsupervised translation network, BalaGAN, specifically designed to tackle the domain imbalance problem. We leverage the latent modalities of the richer domain to turn the image-to-image translation problem, between two imbalanced domains, into a balanced, multi-class, and conditional translation problem, more resembling the style transfer setting. Specifically, we analyze the source domain and learn a decomposition of it into a set of latent modes or classes, without any supervision. This leaves us with a multitude of balanced cross-domain translation tasks, between all pairs of classes, including the target domain. During inference, the trained network takes as input a source image, as well as a reference or style image from one of the modes as a condition, and produces an image which resembles the source on the pixel-wise level, but shares the same mode as the reference. We show that employing modalities within the dataset improves the quality of the translated images, and that BalaGAN outperforms strong baselines of both unconditioned and style-transfer-based image-to-image translation methods, in terms of image quality and diversity.
We present a simple method to reconstruct a high-resolution video from a face-video, where the identity of a person is obscured by pixelization. This concealment method is popular because the viewer can still perceive a human face figure and the overall head motion. However, we show in our experiments that a fairly good approximation of the original video can be reconstructed in a way that compromises anonymity. Our system exploits the simultaneous similarity and small disparity between close-by video frames depicting a human face, and employs a spatial transformation component that learns the alignment between the pixelated frames. Each frame, supported by its aligned surrounding frames, is first encoded, then decoded to a higher resolution. Reconstruction and perceptual losses promote adherence to the ground-truth, and an adversarial loss assists in maintaining domain faithfulness. There is no need for explicit temporal coherency loss as it is maintained implicitly by the alignment of neighboring frames and reconstruction. Although simple, our framework synthesizes high-quality face reconstructions, demonstrating that given the statistical prior of a human face, multiple aligned pixelated frames contain sufficient information to reconstruct a high-quality approximation of the original signal.
The paradigm of image-to-image translation is leveraged for the benefit of sketch stylization via transfer of geometric textural details. Lacking the necessary volumes of data for standard training of translation systems, we advocate for operation at the patch level, where a handful of stylized sketches provide ample mining potential for patches featuring basic geometric primitives. Operating at the patch level necessitates special consideration of full sketch translation, as individual translation of patches with no regard to neighbors is likely to produce visible seams and artifacts at patch borders. Aligned pairs of styled and plain primitives are combined to form input hybrids containing styled elements around the border and plain elements within, and given as input to a seamless translation (ST) generator, whose output patches are expected to reconstruct the fully styled patch. An adversarial addition promotes generalization and robustness to diverse geometries at inference time, forming a simple and effective system for arbitrary sketch stylization, as demonstrated upon a variety of styles and sketches.
Humans regularly interact with their surrounding objects. Such interactions often result in strongly correlated motion between humans and the interacting objects. We thus ask: "Is it possible to infer object properties from skeletal motion alone, even without seeing the interacting object itself?" In this paper, we present a fine-grained action recognition method that learns to infer such latent object properties from human interaction motion alone. This inference allows us to disentangle the motion from the object property and transfer object properties to a given motion. We collected a large number of videos and 3D skeletal motions of the performing actors using an inertial motion capture device. We analyze similar actions and learn subtle differences among them to reveal latent properties of the interacting objects. In particular, we learn to identify the interacting object, by estimating its weight, or its fragility or delicacy. Our results clearly demonstrate that the interaction motions and interacting objects are highly correlated and indeed relative object latent properties can be inferred from the 3D skeleton sequences alone, leading to new synthesis possibilities for human interaction motions. Dataset will be available at http://vcc.szu.edu.cn/research/2020/IT.
In this paper, we introduce a deep learning technique for consolidating and sharp feature generation of point clouds using only the input point cloud itself. Rather than explicitly define a prior that describes typical shape characteristics (i.e., piecewise-smoothness), or a heuristic policy for generating novel sharp points, we opt to learn both using a neural network with shared-weights. Instead of relying on a large collection of manually annotated data, we use the self-supervision present within a single shape, i.e., self-prior, to train the network, and learn the underlying distribution of sharp features specific to the given input point cloud. By learning to map a low-curvature subset of the input point cloud to a disjoint high-curvature subset, the network formalizes the shape-specific characteristics and infers how to generate sharp points. During test time, the network is repeatedly fed a random subset of points from the input and displaces them to generate an arbitrarily large set of novel sharp feature points. The local shared weights are optimized over the entire shape, learning non-local statistics and exploiting the recurrence of local-scale geometries. We demonstrate the ability to generate coherent sets of sharp feature points on a variety of shapes, while eliminating outliers and noise.
We present a generic image-to-image translation framework, Pixel2Style2Pixel (pSp). Our pSp framework is based on a novel encoder network that directly generates a series of style vectors which are fed into a pretrained StyleGAN generator, forming the extended W+ latent space. We first show that our encoder can directly embed real images into W+, with no additional optimization. We further introduce a dedicated identity loss which is shown to achieve improved performance in the reconstruction of an input image. We demonstrate pSp to be a simple architecture that, by leveraging a well-trained, fixed generator network, can be easily applied on a wide-range of image-to-image translation tasks. Solving these tasks through the style representation results in a global approach that does not rely on a local pixel-to-pixel correspondence and further supports multi-modal synthesis via the resampling of styles. Notably, we demonstrate that pSp can be trained to align a face image to a frontal pose without any labeled data, generate multi-modal results for ambiguous tasks such as conditional face generation from segmentation maps, and construct high-resolution images from corresponding low-resolution images.
We present MRGAN, a multi-rooted adversarial network which generates part-disentangled 3D point-cloud shapes without part-based shape supervision. The network fuses multiple branches of tree-structured graph convolution layers which produce point clouds, with learnable constant inputs at the tree roots. Each branch learns to grow a different shape part, offering control over the shape generation at the part level. Our network encourages disentangled generation of semantic parts via two key ingredients: a root-mixing training strategy which helps decorrelate the different branches to facilitate disentanglement, and a set of loss terms designed with part disentanglement and shape semantics in mind. Of these, a novel convexity loss incentivizes the generation of parts that are more convex, as semantic parts tend to be. In addition, a root-dropping loss further ensures that each root seeds a single part, preventing the degeneration or over-growth of the point-producing branches. We evaluate the performance of our network on a number of 3D shape classes, and offer qualitative and quantitative comparisons to previous works and baseline approaches. We demonstrate the controllability offered by our part-disentangled generation through two applications for shape modeling: part mixing and individual part variation, without receiving segmented shapes as input.
Recently, deep generative adversarial networks for image generation have advanced rapidly; yet, only a small amount of research has focused on generative models for irregular structures, particularly meshes. Nonetheless, mesh generation and synthesis remains a fundamental topic in computer graphics. In this work, we propose a novel framework for synthesizing geometric textures. It learns geometric texture statistics from local neighborhoods (i.e., local triangular patches) of a single reference 3D model. It learns deep features on the faces of the input triangulation, which is used to subdivide and generate offsets across multiple scales, without parameterization of the reference or target mesh. Our network displaces mesh vertices in any direction (i.e., in the normal and tangential direction), enabling synthesis of geometric textures, which cannot be expressed by a simple 2D displacement map. Learning and synthesizing on local geometric patches enables a genus-oblivious framework, facilitating texture transfer between shapes of different genus.