High-quality 3D face asset creation remains costly due to reliance on controlled capture setups and manual processing, limiting scalability and diversity. We introduce a fully automated, semantically controllable framework for generating PBR-ready 3D facial assets without requiring dedicated scans. Our pipeline begins with a diffusion-based data synthesis stage, where 2D portrait samples from a pre-trained diffusion model are converted into 44K textured 3D face reconstructions via our proposed geometry recovery and texture normalization algorithm, which aligns arbitrarily shaded outputs into clean albedo space. Using this dataset, we train a disentangled adversarial generator that maps semantic attributes (age, gender, ethnicity) to UV-space geometry and albedo, enabling both direct sampling and continuous latent editing while preserving identity. A refinement stage further produces PBR materials and secondary assets (eyeballs, teeth, gums). The resulting system supports controllable face generation and post-editing in real time and exports directly to standard rendering and animation pipelines. We evaluate each component extensively and provide a web-based interactive interface to showcase practical deployment.

Disentangling data into interpretable and independent factors is critical for controllable generation tasks. With the availability of labeled data, supervision can help enforce the separation of specific factors as expected. However, it is often expensive or even impossible to label every single factor to achieve fully-supervised disentanglement. In this paper, we adopt a general setting where all factors that are hard to label or identify are encapsulated as a single unknown factor. Under this setting, we propose a flexible weakly-supervised multi-factor disentanglement framework DisUnknown, which Distills Unknown factors for enabling multi-conditional generation regarding both labeled and unknown factors. Specifically, a two-stage training approach is adopted to first disentangle the unknown factor with an effective and robust training method, and then train the final generator with the proper disentanglement of all labeled factors utilizing the unknown distillation. To demonstrate the generalization capacity and scalability of our method, we evaluate it on multiple benchmark datasets qualitatively and quantitatively and further apply it to various real-world applications on complicated datasets.

Many problems in computer graphics and computer vision applications involves inferring a rotation from a variety of different forms of inputs. With the increasing use of deep learning, neural networks have been employed to solve such problems. However, the traditional representations for 3D rotations, the quaternions and Euler angles, are found to be problematic for neural networks in practice, producing seemingly unavoidable large estimation errors. Previous researches has identified the discontinuity of the mapping from SO(3) to the quaternions or Euler angles as the source of such errors, and to solve it, embeddings of SO(3) have been proposed as the output representation of rotation estimation networks instead. In this paper, we argue that the argument against quaternions and Euler angles from local discontinuities of the mappings from SO(3) is flawed, and instead provide a different argument from the global topological properties of SO(3) that also establishes the lower bound of maximum error when using quaternions and Euler angles for rotation estimation networks. Extending from this view, we discover that rotation symmetries in the input object causes additional topological problems that even using embeddings of SO(3) as the output representation would not correctly handle. We propose the self-selecting ensemble, a topologically motivated approach, where the network makes multiple predictions and assigns weights to them. We show theoretically and with experiments that our methods can be combined with a wide range of different rotation representations and can handle all kinds of finite symmetries in 3D rotation estimation problems.

We present a novel method to realistically puppeteer and animate a face from a single RGB image using a source video sequence. We begin by fitting a multilinear PCA model to obtain the 3D geometry and a single texture of the target face. In order for the animation to be realistic, however, we need dynamic per-frame textures that capture subtle wrinkles and deformations corresponding to the animated facial expressions. This problem is highly underconstrained, as dynamic textures cannot be obtained directly from a single image. Furthermore, if the target face has a closed mouth, it is not possible to obtain actual images of the mouth interior. To address this issue, we train a Deep Generative Network that can infer realistic per-frame texture deformations, including the mouth interior, of the target identity using the per-frame source textures and the single target texture. By retargeting the PCA expression geometry from the source, as well as using the newly inferred texture, we can both animate the face and perform video face replacement on the source video using the target appearance.