Synthetic Prior for Few-Shot Drivable Head Avatar Inversion

We present SynShot, a novel method for the few-shot inversion of a drivable head avatar based on a synthetic prior. We tackle two major challenges. First, training a controllable 3D generative network requires a large number of diverse sequences, for which pairs of images and high-quality tracked meshes are not always available. Second, state-of-the-art monocular avatar models struggle to generalize to new views and expressions, lacking a strong prior and often overfitting to a specific viewpoint distribution. Inspired by machine learning models trained solely on synthetic data, we propose a method that learns a prior model from a large dataset of synthetic heads with diverse identities, expressions, and viewpoints. With few input images, SynShot fine-tunes the pretrained synthetic prior to bridge the domain gap, modeling a photorealistic head avatar that generalizes to novel expressions and viewpoints. We model the head avatar using 3D Gaussian splatting and a convolutional encoder-decoder that outputs Gaussian parameters in UV texture space. To account for the different modeling complexities over parts of the head (e.g., skin vs hair), we embed the prior with explicit control for upsampling the number of per-part primitives. Compared to state-of-the-art monocular methods that require thousands of real training images, SynShot significantly improves novel view and expression synthesis.

Learning to Stabilize Faces

Nowadays, it is possible to scan faces and automatically register them with high quality. However, the resulting face meshes often need further processing: we need to stabilize them to remove unwanted head movement. Stabilization is important for tasks like game development or movie making which require facial expressions to be cleanly separated from rigid head motion. Since manual stabilization is labor-intensive, there have been attempts to automate it. However, previous methods remain impractical: they either still require some manual input, produce imprecise alignments, rely on dubious heuristics and slow optimization, or assume a temporally ordered input. Instead, we present a new learning-based approach that is simple and fully automatic. We treat stabilization as a regression problem: given two face meshes, our network directly predicts the rigid transform between them that brings their skulls into alignment. We generate synthetic training data using a 3D Morphable Model (3DMM), exploiting the fact that 3DMM parameters separate skull motion from facial skin motion. Through extensive experiments we show that our approach outperforms the state-of-the-art both quantitatively and qualitatively on the tasks of stabilizing discrete sets of facial expressions as well as dynamic facial performances. Furthermore, we provide an ablation study detailing the design choices and best practices to help others adopt our approach for their own uses. Supplementary videos can be found on the project webpage syntec-research.github.io/FaceStab.

OFER: Occluded Face Expression Reconstruction

Reconstructing 3D face models from a single image is an inherently ill-posed problem, which becomes even more challenging in the presence of occlusions. In addition to fewer available observations, occlusions introduce an extra source of ambiguity, where multiple reconstructions can be equally valid. Despite the ubiquity of the problem, very few methods address its multi-hypothesis nature. In this paper we introduce OFER, a novel approach for single image 3D face reconstruction that can generate plausible, diverse, and expressive 3D faces, even under strong occlusions. Specifically, we train two diffusion models to generate the shape and expression coefficients of a face parametric model, conditioned on the input image. This approach captures the multi-modal nature of the problem, generating a distribution of solutions as output. Although this addresses the ambiguity problem, the challenge remains to pick the best matching shape to ensure consistency across diverse expressions. To achieve this, we propose a novel ranking mechanism that sorts the outputs of the shape diffusion network based on the predicted shape accuracy scores to select the best match. We evaluate our method using standard benchmarks and introduce CO-545, a new protocol and dataset designed to assess the accuracy of expressive faces under occlusion. Our results show improved performance over occlusion-based methods, with added ability to generate multiple expressions for a given image.

Gaussian Eigen Models for Human Heads

We present personalized Gaussian Eigen Models (GEMs) for human heads, a novel method that compresses dynamic 3D Gaussians into low-dimensional linear spaces. Our approach is inspired by the seminal work of Blanz and Vetter, where a mesh-based 3D morphable model (3DMM) is constructed from registered meshes. Based on dynamic 3D Gaussians, we create a lower-dimensional representation of primitives that applies to most 3DGS head avatars. Specifically, we propose a universal method to distill the appearance of a mesh-controlled UNet Gaussian avatar using an ensemble of linear eigenbasis. We replace heavy CNN-based architectures with a single linear layer improving speed and enabling a range of real-time downstream applications. To create a particular facial expression, one simply needs to perform a dot product between the eigen coefficients and the distilled basis. This efficient method removes the requirement for an input mesh during testing, enhancing simplicity and speed in expression generation. This process is highly efficient and supports real-time rendering on everyday devices, leveraging the effectiveness of standard Gaussian Splatting. In addition, we demonstrate how the GEM can be controlled using a ResNet-based regression architecture. We show and compare self-reenactment and cross-person reenactment to state-of-the-art 3D avatar methods, demonstrating higher quality and better control. A real-time demo showcases the applicability of the GEM representation.

3D Facial Expressions through Analysis-by-Neural-Synthesis

While existing methods for 3D face reconstruction from in-the-wild images excel at recovering the overall face shape, they commonly miss subtle, extreme, asymmetric, or rarely observed expressions. We improve upon these methods with SMIRK (Spatial Modeling for Image-based Reconstruction of Kinesics), which faithfully reconstructs expressive 3D faces from images. We identify two key limitations in existing methods: shortcomings in their self-supervised training formulation, and a lack of expression diversity in the training images. For training, most methods employ differentiable rendering to compare a predicted face mesh with the input image, along with a plethora of additional loss functions. This differentiable rendering loss not only has to provide supervision to optimize for 3D face geometry, camera, albedo, and lighting, which is an ill-posed optimization problem, but the domain gap between rendering and input image further hinders the learning process. Instead, SMIRK replaces the differentiable rendering with a neural rendering module that, given the rendered predicted mesh geometry, and sparsely sampled pixels of the input image, generates a face image. As the neural rendering gets color information from sampled image pixels, supervising with neural rendering-based reconstruction loss can focus solely on the geometry. Further, it enables us to generate images of the input identity with varying expressions while training. These are then utilized as input to the reconstruction model and used as supervision with ground truth geometry. This effectively augments the training data and enhances the generalization for diverse expressions. Our qualitative, quantitative and particularly our perceptual evaluations demonstrate that SMIRK achieves the new state-of-the art performance on accurate expression reconstruction. Project webpage: https://georgeretsi.github.io/smirk/.

Emotional Speech-driven 3D Body Animation via Disentangled Latent Diffusion

Existing methods for synthesizing 3D human gestures from speech have shown promising results, but they do not explicitly model the impact of emotions on the generated gestures. Instead, these methods directly output animations from speech without control over the expressed emotion. To address this limitation, we present AMUSE, an emotional speech-driven body animation model based on latent diffusion. Our observation is that content (i.e., gestures related to speech rhythm and word utterances), emotion, and personal style are separable. To account for this, AMUSE maps the driving audio to three disentangled latent vectors: one for content, one for emotion, and one for personal style. A latent diffusion model, trained to generate gesture motion sequences, is then conditioned on these latent vectors. Once trained, AMUSE synthesizes 3D human gestures directly from speech with control over the expressed emotions and style by combining the content from the driving speech with the emotion and style of another speech sequence. Randomly sampling the noise of the diffusion model further generates variations of the gesture with the same emotional expressivity. Qualitative, quantitative, and perceptual evaluations demonstrate that AMUSE outputs realistic gesture sequences. Compared to the state of the art, the generated gestures are better synchronized with the speech content and better represent the emotion expressed by the input speech. Our project website is amuse.is.tue.mpg.de.

Learning Disentangled Avatars with Hybrid 3D Representations

Tremendous efforts have been made to learn animatable and photorealistic human avatars. Towards this end, both explicit and implicit 3D representations are heavily studied for a holistic modeling and capture of the whole human (e.g., body, clothing, face and hair), but neither representation is an optimal choice in terms of representation efficacy since different parts of the human avatar have different modeling desiderata. For example, meshes are generally not suitable for modeling clothing and hair. Motivated by this, we present Disentangled Avatars~(DELTA), which models humans with hybrid explicit-implicit 3D representations. DELTA takes a monocular RGB video as input, and produces a human avatar with separate body and clothing/hair layers. Specifically, we demonstrate two important applications for DELTA. For the first one, we consider the disentanglement of the human body and clothing and in the second, we disentangle the face and hair. To do so, DELTA represents the body or face with an explicit mesh-based parametric 3D model and the clothing or hair with an implicit neural radiance field. To make this possible, we design an end-to-end differentiable renderer that integrates meshes into volumetric rendering, enabling DELTA to learn directly from monocular videos without any 3D supervision. Finally, we show that how these two applications can be easily combined to model full-body avatars, such that the hair, face, body and clothing can be fully disentangled yet jointly rendered. Such a disentanglement enables hair and clothing transfer to arbitrary body shapes. We empirically validate the effectiveness of DELTA's disentanglement by demonstrating its promising performance on disentangled reconstruction, virtual clothing try-on and hairstyle transfer. To facilitate future research, we also release an open-sourced pipeline for the study of hybrid human avatar modeling.

SCULPT: Shape-Conditioned Unpaired Learning of Pose-dependent Clothed and Textured Human Meshes

We present SCULPT, a novel 3D generative model for clothed and textured 3D meshes of humans. Specifically, we devise a deep neural network that learns to represent the geometry and appearance distribution of clothed human bodies. Training such a model is challenging, as datasets of textured 3D meshes for humans are limited in size and accessibility. Our key observation is that there exist medium-sized 3D scan datasets like CAPE, as well as large-scale 2D image datasets of clothed humans and multiple appearances can be mapped to a single geometry. To effectively learn from the two data modalities, we propose an unpaired learning procedure for pose-dependent clothed and textured human meshes. Specifically, we learn a pose-dependent geometry space from 3D scan data. We represent this as per vertex displacements w.r.t. the SMPL model. Next, we train a geometry conditioned texture generator in an unsupervised way using the 2D image data. We use intermediate activations of the learned geometry model to condition our texture generator. To alleviate entanglement between pose and clothing type, and pose and clothing appearance, we condition both the texture and geometry generators with attribute labels such as clothing types for the geometry, and clothing colors for the texture generator. We automatically generated these conditioning labels for the 2D images based on the visual question answering model BLIP and CLIP. We validate our method on the SCULPT dataset, and compare to state-of-the-art 3D generative models for clothed human bodies. We will release the codebase for research purposes.

Emotional Speech-Driven Animation with Content-Emotion Disentanglement

To be widely adopted, 3D facial avatars need to be animated easily, realistically, and directly, from speech signals. While the best recent methods generate 3D animations that are synchronized with the input audio, they largely ignore the impact of emotions on facial expressions. Instead, their focus is on modeling the correlations between speech and facial motion, resulting in animations that are unemotional or do not match the input emotion. We observe that there are two contributing factors resulting in facial animation - the speech and the emotion. We exploit these insights in EMOTE (Expressive Model Optimized for Talking with Emotion), which generates 3D talking head avatars that maintain lip sync while enabling explicit control over the expression of emotion. Due to the absence of high-quality aligned emotional 3D face datasets with speech, EMOTE is trained from an emotional video dataset (i.e., MEAD). To achieve this, we match speech-content between generated sequences and target videos differently from emotion content. Specifically, we train EMOTE with additional supervision in the form of a lip-reading objective to preserve the speech-dependent content (spatially local and high temporal frequency), while utilizing emotion supervision on a sequence-level (spatially global and low frequency). Furthermore, we employ a content-emotion exchange mechanism in order to supervise different emotion on the same audio, while maintaining the lip motion synchronized with the speech. To employ deep perceptual losses without getting undesirable artifacts, we devise a motion prior in form of a temporal VAE. Extensive qualitative, quantitative, and perceptual evaluations demonstrate that EMOTE produces state-of-the-art speech-driven facial animations, with lip sync on par with the best methods while offering additional, high-quality emotional control.

Instant Multi-View Head Capture through Learnable Registration

Existing methods for capturing datasets of 3D heads in dense semantic correspondence are slow, and commonly address the problem in two separate steps; multi-view stereo (MVS) reconstruction followed by non-rigid registration. To simplify this process, we introduce TEMPEH (Towards Estimation of 3D Meshes from Performances of Expressive Heads) to directly infer 3D heads in dense correspondence from calibrated multi-view images. Registering datasets of 3D scans typically requires manual parameter tuning to find the right balance between accurately fitting the scans surfaces and being robust to scanning noise and outliers. Instead, we propose to jointly register a 3D head dataset while training TEMPEH. Specifically, during training we minimize a geometric loss commonly used for surface registration, effectively leveraging TEMPEH as a regularizer. Our multi-view head inference builds on a volumetric feature representation that samples and fuses features from each view using camera calibration information. To account for partial occlusions and a large capture volume that enables head movements, we use view- and surface-aware feature fusion, and a spatial transformer-based head localization module, respectively. We use raw MVS scans as supervision during training, but, once trained, TEMPEH directly predicts 3D heads in dense correspondence without requiring scans. Predicting one head takes about 0.3 seconds with a median reconstruction error of 0.26 mm, 64% lower than the current state-of-the-art. This enables the efficient capture of large datasets containing multiple people and diverse facial motions. Code, model, and data are publicly available at https://tempeh.is.tue.mpg.de.