Can a generative model be trained to produce images from a specific domain, guided by a text prompt only, without seeing any image? In other words: can an image generator be trained blindly? Leveraging the semantic power of large scale Contrastive-Language-Image-Pre-training (CLIP) models, we present a text-driven method that allows shifting a generative model to new domains, without having to collect even a single image from those domains. We show that through natural language prompts and a few minutes of training, our method can adapt a generator across a multitude of domains characterized by diverse styles and shapes. Notably, many of these modifications would be difficult or outright impossible to reach with existing methods. We conduct an extensive set of experiments and comparisons across a wide range of domains. These demonstrate the effectiveness of our approach and show that our shifted models maintain the latent-space properties that make generative models appealing for downstream tasks.
We propose a novel method for solving regression tasks using few-shot or weak supervision. At the core of our method is the fundamental observation that GANs are incredibly successful at encoding semantic information within their latent space, even in a completely unsupervised setting. For modern generative frameworks, this semantic encoding manifests as smooth, linear directions which affect image attributes in a disentangled manner. These directions have been widely used in GAN-based image editing. We show that such directions are not only linear, but that the magnitude of change induced on the respective attribute is approximately linear with respect to the distance traveled along them. By leveraging this observation, our method turns a pre-trained GAN into a regression model, using as few as two labeled samples. This enables solving regression tasks on datasets and attributes which are difficult to produce quality supervision for. Additionally, we show that the same latent-distances can be used to sort collections of images by the strength of given attributes, even in the absence of explicit supervision. Extensive experimental evaluations demonstrate that our method can be applied across a wide range of domains, leverage multiple latent direction discovery frameworks, and achieve state-of-the-art results in few-shot and low-supervision settings, even when compared to methods designed to tackle a single task.
We present StyleFusion, a new mapping architecture for StyleGAN, which takes as input a number of latent codes and fuses them into a single style code. Inserting the resulting style code into a pre-trained StyleGAN generator results in a single harmonized image in which each semantic region is controlled by one of the input latent codes. Effectively, StyleFusion yields a disentangled representation of the image, providing fine-grained control over each region of the generated image. Moreover, to help facilitate global control over the generated image, a special input latent code is incorporated into the fused representation. StyleFusion operates in a hierarchical manner, where each level is tasked with learning to disentangle a pair of image regions (e.g., the car body and wheels). The resulting learned disentanglement allows one to modify both local, fine-grained semantics (e.g., facial features) as well as more global features (e.g., pose and background), providing improved flexibility in the synthesis process. As a natural extension, StyleFusion enables one to perform semantically-aware cross-image mixing of regions that are not necessarily aligned. Finally, we demonstrate how StyleFusion can be paired with existing editing techniques to more faithfully constrain the edit to the user's region of interest.
The task of unsupervised motion retargeting in videos has seen substantial advancements through the use of deep neural networks. While early works concentrated on specific object priors such as a human face or body, recent work considered the unsupervised case. When the source and target videos, however, are of different shapes, current methods fail. To alleviate this problem, we introduce JOKR - a JOint Keypoint Representation that captures the motion common to both the source and target videos, without requiring any object prior or data collection. By employing a domain confusion term, we enforce the unsupervised keypoint representations of both videos to be indistinguishable. This encourages disentanglement between the parts of the motion that are common to the two domains, and their distinctive appearance and motion, enabling the generation of videos that capture the motion of the one while depicting the style of the other. To enable cases where the objects are of different proportions or orientations, we apply a learned affine transformation between the JOKRs. This augments the representation to be affine invariant, and in practice broadens the variety of possible retargeting pairs. This geometry-driven representation enables further intuitive control, such as temporal coherence and manual editing. Through comprehensive experimentation, we demonstrate the applicability of our method to different challenging cross-domain video pairs. We evaluate our method both qualitatively and quantitatively, and demonstrate that our method handles various cross-domain scenarios, such as different animals, different flowers, and humans. We also demonstrate superior temporal coherency and visual quality compared to state-of-the-art alternatives, through statistical metrics and a user study. Source code and videos can be found at https://rmokady.github.io/JOKR/ .
Recently, a surge of advanced facial editing techniques have been proposed that leverage the generative power of a pre-trained StyleGAN. To successfully edit an image this way, one must first project (or invert) the image into the pre-trained generator's domain. As it turns out, however, StyleGAN's latent space induces an inherent tradeoff between distortion and editability, i.e. between maintaining the original appearance and convincingly altering some of its attributes. Practically, this means it is still challenging to apply ID-preserving facial latent-space editing to faces which are out of the generator's domain. In this paper, we present an approach to bridge this gap. Our technique slightly alters the generator, so that an out-of-domain image is faithfully mapped into an in-domain latent code. The key idea is pivotal tuning - a brief training process that preserves the editing quality of an in-domain latent region, while changing its portrayed identity and appearance. In Pivotal Tuning Inversion (PTI), an initial inverted latent code serves as a pivot, around which the generator is fined-tuned. At the same time, a regularization term keeps nearby identities intact, to locally contain the effect. This surgical training process ends up altering appearance features that represent mostly identity, without affecting editing capabilities. We validate our technique through inversion and editing metrics, and show preferable scores to state-of-the-art methods. We further qualitatively demonstrate our technique by applying advanced edits (such as pose, age, or expression) to numerous images of well-known and recognizable identities. Finally, we demonstrate resilience to harder cases, including heavy make-up, elaborate hairstyles and/or headwear, which otherwise could not have been successfully inverted and edited by state-of-the-art methods.
Synthesizing novel views of dynamic humans from stationary monocular cameras is a popular scenario. This is particularly attractive as it does not require static scenes, controlled environments, or specialized hardware. In contrast to techniques that exploit multi-view observations to constrain the modeling, given a single fixed viewpoint only, the problem of modeling the dynamic scene is significantly more under-constrained and ill-posed. In this paper, we introduce Neural Motion Consensus Flow (MoCo-Flow), a representation that models the dynamic scene using a 4D continuous time-variant function. The proposed representation is learned by an optimization which models a dynamic scene that minimizes the error of rendering all observation images. At the heart of our work lies a novel optimization formulation, which is constrained by a motion consensus regularization on the motion flow. We extensively evaluate MoCo-Flow on several datasets that contain human motions of varying complexity, and compare, both qualitatively and quantitatively, to several baseline methods and variants of our methods. Pretrained model, code, and data will be released for research purposes upon paper acceptance.
Rigid registration of partial observations is a fundamental problem in various applied fields. In computer graphics, special attention has been given to the registration between two partial point clouds generated by scanning devices. State-of-the-art registration techniques still struggle when the overlap region between the two point clouds is small, and completely fail if there is no overlap between the scan pairs. In this paper, we present a learning-based technique that alleviates this problem, and allows registration between point clouds, presented in arbitrary poses, and having little or even no overlap, a setting that has been referred to as tele-registration. Our technique is based on a novel neural network design that learns a prior of a class of shapes and can complete a partial shape. The key idea is combining the registration and completion tasks in a way that reinforces each other. In particular, we simultaneously train the registration network and completion network using two coupled flows, one that register-and-complete, and one that complete-and-register, and encourage the two flows to produce a consistent result. We show that, compared with each separate flow, this two-flow training leads to robust and reliable tele-registration, and hence to a better point cloud prediction that completes the registered scans. It is also worth mentioning that each of the components in our neural network outperforms state-of-the-art methods in both completion and registration. We further analyze our network with several ablation studies and demonstrate its performance on a large number of partial point clouds, both synthetic and real-world, that have only small or no overlap.
We present a technique for rendering point clouds using a neural network. Existing point rendering techniques either use splatting, or first reconstruct a surface mesh that can then be rendered. Both of these techniques require solving for global point normal orientation, which is a challenging problem on its own. Furthermore, splatting techniques result in holes and overlaps, whereas mesh reconstruction is particularly challenging, especially in the cases of thin surfaces and sheets. We cast the rendering problem as a conditional image-to-image translation problem. In our formulation, Z2P, i.e., depth-augmented point features as viewed from target camera view, are directly translated by a neural network to rendered images, conditioned on control variables (e.g., color, light). We avoid inevitable issues with splatting (i.e., holes and overlaps), and bypass solving the notoriously challenging surface reconstruction problem or estimating oriented normals. Yet, our approach results in a rendered image as if a surface mesh was reconstructed. We demonstrate that our framework produces a plausible image, and can effectively handle noise, non-uniform sampling, thin surfaces / sheets, and is fast.
The increasing availability of video recordings made by multiple cameras has offered new means for mitigating occlusion and depth ambiguities in pose and motion reconstruction methods. Yet, multi-view algorithms strongly depend on camera parameters, in particular, the relative positions among the cameras. Such dependency becomes a hurdle once shifting to dynamic capture in uncontrolled settings. We introduce FLEX (Free muLti-view rEconstruXion), an end-to-end parameter-free multi-view model. FLEX is parameter-free in the sense that it does not require any camera parameters, neither intrinsic nor extrinsic. Our key idea is that the 3D angles between skeletal parts, as well as bone lengths, are invariant to the camera position. Hence, learning 3D rotations and bone lengths rather than locations allows predicting common values for all camera views. Our network takes multiple video streams, learns fused deep features through a novel multi-view fusion layer, and reconstructs a single consistent skeleton with temporally coherent joint rotations. We demonstrate quantitative and qualitative results on the Human3.6M and KTH Multi-view Football II datasets. We compare our model to state-of-the-art methods that are not parameter-free and show that in the absence of camera parameters, we outperform them by a large margin while obtaining comparable results when camera parameters are available. Code, trained models, video demonstration, and additional materials will be available on our project page.
Establishing a consistent normal orientation for point clouds is a notoriously difficult problem in geometry processing, requiring attention to both local and global shape characteristics. The normal direction of a point is a function of the local surface neighborhood; yet, point clouds do not disclose the full underlying surface structure. Even assuming known geodesic proximity, calculating a consistent normal orientation requires the global context. In this work, we introduce a novel approach for establishing a globally consistent normal orientation for point clouds. Our solution separates the local and global components into two different sub-problems. In the local phase, we train a neural network to learn a coherent normal direction per patch (i.e., consistently oriented normals within a single patch). In the global phase, we propagate the orientation across all coherent patches using a dipole propagation. Our dipole propagation decides to orient each patch using the electric field defined by all previously orientated patches. This gives rise to a global propagation that is stable, as well as being robust to nearby surfaces, holes, sharp features and noise.