Abstract Facial animation is one of the most labor‐intensive aspects of animation and VFX, as traditional rigging consumes weeks of expert time and forces animators to spend countless hours manipulating hundreds of controls to achieve varied expressions. This technical complexity creates a barrier between artistic vision and execution, limiting creative exploration and iteration. In this paper, we introduce CANRig , a fully automated neural facial rigging approach that simplifies the process of creating and editing facial poses by benefiting from global correlations learned from data. Unlike existing neural face models that either sacrifice local control or demand extensive manual region setup, our method introduces continuous local control through a novel conditioning mechanism that operates on a variable region. By modeling deformation as cross‐attention between control handles and mesh vertices—modulated by a user‐defined region—we enable seamless transitions from precise local adjustments to broad global changes. We further expand our method with a shape‐preserving workflow that enables iterative edits, guaranteeing that changes remain untouched even as controls are reconfigured. Our method delivers the best of both worlds: the automation and naturalness of neural methods with the granular control that professional animators demand, and we demonstrate its effectiveness across multiple applications in both animation and high‐end visual effects pipelines.

Abstract Human motion data is inherently rich and complex, containing both semantic content and subtle stylistic features that are challenging to model. We propose a novel method for effective disentanglement of the style and content in human motion data to facilitate style transfer. Our approach is guided by the insight that content corresponds to coarse motion attributes while style captures the finer, expressive details. To model this hierarchy, we employ Residual Vector Quantized Variational Autoencoders (RVQ‐VAEs) to learn a coarse‐to‐fine representation of motion. We further enhance the disentanglement by integrating codebook learning with contrastive learning and a novel information leakage loss to organize the content and the style across different codebooks. We harness this disentangled representation using our simple and effective inference‐time technique Quantized Code Swapping , which enables motion style transfer without requiring any fine‐tuning for unseen styles. Our framework demonstrates strong versatility across multiple inference applications, including style transfer, style removal, and motion blending.

Realistic animal motion for virtual production is typically obtained either through motion capture of highly trained performers who accurately mimic animal behavior, or by retargeting ordinary human motion using complex control setups. Both approaches are challenging and often fail to fully reproduce the nuances of natural animal motion, motivating data-driven alternatives. We present an automatic human-to-quadruped puppeteering framework that produces plausible and controllable quadruped motions from ordinary human motion data. Our approach employs a two-stage generative diffusion model trained purely on quadruped motion data. By introducing a structured conditioning and inpainting strategy, our method supports a wide range of actions, including walking, running, jumping, sitting, and lying. Furthermore, we enable fine-grained intuitive control of the quadruped motion such as head movement control and individual limb puppeteering. Experimental results demonstrate improved motion realism and controllability compared to existing retargeting approaches, highlighting the effectiveness of our framework as a tool for animation and virtual production applications.

Creating high-quality character animation remains an intricate and cumbersome process that requires skill, training, and craftsmanship to master. Recently, diffusion models have unlocked the ability to generate diverse movements from high-level condition signals such as text. For artist-friendly control, motion diffusion leveraging Bézier curves have been shown to allow precise joint-level conditioning. Yet, these works have been limited to joints at a fixed temporal stride, while animators require more temporal flexibility when keyframing or manipulating tangents to achieve animation principles such as easing in & out. In this work, we introduce a new Implicit Bézier Motion Model (IBMM), which during training is exposed to all possible configurations of control points, enabling control at arbitrary timings. This allows both precise and sparse joint-level control, anywhere in time and for any joint. In addition, we introduce a new quantitative measure of ease-in and -out, which leads to a novel condition over the motion generation process to reflect this artistic principle.

While conditional diffusion models are known to have good coverage of the data distribution, they still face limitations in output diversity, particularly when sampled with a high classifier-free guidance scale for optimal image quality or when trained on small datasets. We attribute this problem to the role of the conditioning signal in inference and offer an improved sampling strategy for diffusion models that can increase generation diversity, especially at high guidance scales, with minimal loss of sample quality. Our sampling strategy anneals the conditioning signal by adding scheduled, monotonically decreasing Gaussian noise to the conditioning vector during inference to balance diversity and condition alignment. Our Condition-Annealed Diffusion Sampler (CADS) can be used with any pretrained model and sampling algorithm, and we show that it boosts the diversity of diffusion models in various conditional generation tasks. Further, using an existing pretrained diffusion model, CADS achieves a new state-of-the-art FID of 1.70 and 2.31 for class-conditional ImageNet generation at 256$\times$256 and 512$\times$512 respectively.

Animation is a challenging and time-consuming process where animators must manipulate hundreds of controls over space and time to create compelling motions. Recent advances in motion diffusion models have shown impressive results for general motion generation and hold the potential to reduce the number of controls manipulated by animators to achieve high quality results. However, these models are limited by their inability to match sparse constraints precisely, preventing frame-level joint control required by artists. Additionally, recent models are trained for specific characters, preventing reuse, and are incompatible for characters with only a small datasets available. To tackle these shortcomings, we propose a novel factorization of motion between a character-agnostic Bézier Motion Model (BMM), which can be trained on a large motion dataset, followed by a character-specific posing model, trainable on a much smaller pose dataset, that enables reuse across many characters. BMM provides accuracy for meeting sparse joint-level constraints by working in a reduced space of Bézier curves that better aligns the condition signal with the prediction space of our model. Additionally, the Bézier curves offer animators an intuitive interface compatible with existing authoring software. Through quantitative and qualitative comparisons, we show the effectiveness of our factorization and parametric subspace, enabling user control with higher fidelity.

Advances in latent diffusion models (LDMs) have revolutionized high-resolution image generation, but the design space of the autoencoder that is central to these systems remains underexplored. In this paper, we introduce LiteVAE, a new autoencoder design for LDMs, which leverages the 2D discrete wavelet transform to enhance scalability and computational efficiency over standard variational autoencoders (VAEs) with no sacrifice in output quality. We investigate the training methodologies and the decoder architecture of LiteVAE and propose several enhancements that improve the training dynamics and reconstruction quality. Our base LiteVAE model matches the quality of the established VAEs in current LDMs with a six-fold reduction in encoder parameters, leading to faster training and lower GPU memory requirements, while our larger model outperforms VAEs of comparable complexity across all evaluated metrics (rFID, LPIPS, PSNR, and SSIM).

Authoring 3D motions is a laborious process that requires manipulating and coordinating many control handles over time. Neural motion representations learned from large motion datasets have recently shown impressive capabilities in many motion completion tasks. However, current methods are not designed for interactive motion authoring workflows. The reasons being their requirement of a dense context of full poses, which takes considerable time to author, as well as their lack of joint-level controls for refinement. In this paper, we introduce a Neural Motion Rig called SKEL-Betweener, tailored to interactive motion authoring. SKEL-Betweener is able to generate long motion sequences from two poses only, and enables intermediate motion authoring via neural motion curves---intuitive joint-level controls for positions and orientations. Through user evaluations, we demonstrate the effectiveness of our Neural Motion Rig for efficiently creating and editing motions.

Posing a 3D character for film or game is an iterative and laborious process where many control handles (e.g. joints) need to be manipulated to achieve a compelling result. Neural Inverse Kinematics (IK) is a new type of IK that enables sparse control over a 3D character pose, and leverages full body correlations to complete the un-manipulated joints of the body. While neural IK is promising, current methods are not designed to preserve previous edits in posing workflows. Current models generate a single pose from the handles only—regardless of what was there previously—making it difficult to preserve any variations and hindering tasks such as pose and motion editing.

Building intelligent agents that can believably interact with humans is a difficult yet important task in a host of applications, including therapy, education, and entertainment. We submit that in order to enhance believability, the agent's affective state should be accurately modeled and should realistically influence the agent's behavior. We propose a computational model of affect which incorporates an empirically-based interplay between its various affective components - personality, motivation, emotion, and mood. Further, our model captures a number of salient mechanisms that are observable in humans and that influence the agent's behavior. We are therefore hopeful that our model will facilitate more engaging and meaningful human-agent interactions. We evaluate our model and illustrate its efficacy, as well as the importance of the different components in the model and their interplay.