DiffPoseTalk: Speech-Driven Stylistic 3D Facial Animation and Head Pose Generation via Diffusion Models

TL;DR

DiffPoseTalk tackles the challenge of generating diverse, stylistic 3D facial animations driven by speech by casting the problem as a diffusion-based generation task with explicit style and head-pose control. It combines a HuBERT-based speech encoder, a FLAME 3DMM representation, a transformer denoiser, and a speaking style encoder trained with contrastive learning, enabling classifier-free guidance during sampling. Key contributions include (1) a style-controllable diffusion framework that accepts arbitrary reference-style inputs, (2) a windowed training and incremental CFG strategy to stabilize long sequences and balance fidelity with diversity, and (3) a new Talking Face with Head Poses dataset (TFHP) for robust real-world training. Empirical results show state-of-the-art lip synchronization, head pose alignment, and expressive style coverage, with a user study confirming perceptual improvements and practical impact for avatar animation and related applications. $N=500$ steps are used in the diffusion process, and the system operates on FLAME-based 3DMM parameters to enable downstream integration.

Abstract

The generation of stylistic 3D facial animations driven by speech presents a significant challenge as it requires learning a many-to-many mapping between speech, style, and the corresponding natural facial motion. However, existing methods either employ a deterministic model for speech-to-motion mapping or encode the style using a one-hot encoding scheme. Notably, the one-hot encoding approach fails to capture the complexity of the style and thus limits generalization ability. In this paper, we propose DiffPoseTalk, a generative framework based on the diffusion model combined with a style encoder that extracts style embeddings from short reference videos. During inference, we employ classifier-free guidance to guide the generation process based on the speech and style. In particular, our style includes the generation of head poses, thereby enhancing user perception. Additionally, we address the shortage of scanned 3D talking face data by training our model on reconstructed 3DMM parameters from a high-quality, in-the-wild audio-visual dataset. Extensive experiments and user study demonstrate that our approach outperforms state-of-the-art methods. The code and dataset are at https://diffposetalk.github.io .

Figures (3)

• Figure 1: (Left) Transformer-based denoising network. We employ a windowing strategy to generate speech-driven 3D facial animations for inputs of arbitrary length. HuBERT-encoded speech features ${\bm{A}}_{-T_p:T_w}$, prior clean motion parameters ${\bm{X}}_{-T_p:0}^{0}$, current noisy motion parameters ${\bm{X}}_{0:T_w}^{n}$, shape parameters $\bm \beta$, style feature ${\bm{s}}$, and the diffusion timestep $n$ are fed into the transformer decoder. The decoder then predicts clean motion parameters $\hat{{\bm{X}}}_{-T_p:T_w}^{0}$, which are renoised to ${\bm{X}}_{0:T_w}^{n-1}$ for the next denoising iteration. (Right) The speaking style encoder. The style feature ${\bm{s}}$ can be extracted from a sequence of motion parameters ${\bm{X}}_{0:T}$ using a transformer encoder.
• Figure 2: Qualitative comparison with the state of the arts (w/o head pose prediction). Results of different identities are split by dashed lines.
• Figure 3: Qualitative comparison with the state of the arts (w/ head pose prediction). Results of different identities are split by dashed lines. The ">" indicates stress in speech.