[Paper Review] Pose-Controllable Talking Face Generation by Implicitly Modularized Audio-Visual Representation

Hang Zhou, Yasheng Sun+4 more|arXiv (Cornell University)|Apr 22, 2021

Face recognition and analysis77 references24 citations

TL;DR

The paper presents PC-AVS, a pose-controllable, audio-driven talking face generation framework that uses an implicit low-dimensional pose code and modularized audio-visual representations to enable pose control from a separate pose source video without relying on structural intermediate representations.

ABSTRACT

While accurate lip synchronization has been achieved for arbitrary-subject audio-driven talking face generation, the problem of how to efficiently drive the head pose remains. Previous methods rely on pre-estimated structural information such as landmarks and 3D parameters, aiming to generate personalized rhythmic movements. However, the inaccuracy of such estimated information under extreme conditions would lead to degradation problems. In this paper, we propose a clean yet effective framework to generate pose-controllable talking faces. We operate on raw face images, using only a single photo as an identity reference. The key is to modularize audio-visual representations by devising an implicit low-dimension pose code. Substantially, both speech content and head pose information lie in a joint non-identity embedding space. While speech content information can be defined by learning the intrinsic synchronization between audio-visual modalities, we identify that a pose code will be complementarily learned in a modulated convolution-based reconstruction framework. Extensive experiments show that our method generates accurately lip-synced talking faces whose poses are controllable by other videos. Moreover, our model has multiple advanced capabilities including extreme view robustness and talking face frontalization. Code, models, and demo videos are available at https://hangz-nju-cuhk.github.io/projects/PC-AVS.

Motivation & Objective

• Motivate free head-pose control in audio-driven talking face generation.
• Eliminate dependence on explicit 2D/3D structural intermediates by learning a low-dimension pose code.
• Disentangle identity, speech content, and pose in an implicit, data-driven framework.
• Achieve robust lip synchronization while allowing pose transfer from a separate video source.

Proposed method

• Identify a non-identity feature space via targeted data augmentation on target frames to capture pose-related variations.
• Modularize audio-visual information into three spaces: speech content, head pose, and identity.
• Learn speech content with audio-visual synchronization using contrastive learning (InfoNCE) between visual and audio features.
• Define an implicit 12-D pose code derived from a minimal 3D pose prior and map non-identity features to this pose space.
• Use a modulated-convolution-based generator where convolution weights are dynamically scaled by a concatenated latent code from identity, speech content, and pose features.
• Train with a combination of adversarial, reconstruction (L1), perceptual (VGG), and cross-modal contrastive losses to enforce the modularization and pose alignment.
• Infer poses during generation by driving mouth shapes with the identity and speech content spaces while controlling head pose via a pose source video.

Experimental results

Research questions

• RQ1 Can pose be controlled in talking-face generation without explicit 3D pose estimation or landmarks?
• RQ2 Does a low-dimensional, implicit pose code suffice to modulate head movements while preserving lip synchronization?
• RQ3 Can pose from one video source be transferred to drive another identity’s talking face driven by audio?
• RQ4 Does modularizing identity, speech content, and pose improve lip-synchronization and robustness under extreme viewpoints?

Key findings

• The method achieves accurate lip synchronization while enabling pose control from a pose source video.
• On LRW, PC-AVS matches or surpasses baselines in lip synchronization and image quality; on VoxCeleb2, it shows strong lip-sync and pose control performance.
• The approach exhibits extreme-view robustness and can frontalize faces by setting the pose code to zero.
• Ablation studies show the importance of the contrastive audio-visual synchronization loss and the 12-dimensional pose code for successful pose control.
• User studies indicate higher head movement naturalness and video realness for PC-AVS compared to several baselines.