Audio2Face-3D-v3.0 generates 3D facial animations from audio inputs, for use in applications such as video conferencing, virtual reality, and digital content creation.
This model is ready for commercial/non-commercial use. For source code, documentation, helper scripts, packaged builds, and links to all components in the Audio2Face-3D technology stack, visit the Audio2Face-3D GitHub repository
Audio2Face-3D-v3.0 is designed for developers and researchers working on audio-driven animation and emotion detection applications, such as virtual assistants, chatbots, and affective computing systems.
Architecture Type: Transformer, Diffusion Network Architecture: Hubert Number of model parameters: 1.80x10^8
Input:
Input Type(s): Audio Input Format: Array of float Input Parameters: One-Dimensional (1D) Other Properties Related to Input: All audio is resampled to 16KHz
Output:
Output Type(s): Facial motion Output Format: Array of float Output Parameters: Two-Dimensional (2D) Other Properties Related to Output: Facial motion on skin, tongue, jaw, and eyeballs
Our AI models are designed and/or optimized to run on NVIDIA GPU-accelerated systems. By leveraging NVIDIA’s hardware (e.g. GPU cores) and software frameworks (e.g., CUDA libraries), the model achieves faster training and inference times compared to CPU-only solutions.
The integration of foundation and fine-tuned models into AI systems requires additional testing using use-case-specific data to ensure safe and effective deployment. Following the V-model methodology, iterative testing and validation at both unit and system levels are essential to mitigate risks, meet technical and functional requirements, and ensure compliance with safety and ethical standards before deployment.
This AI model can be embedded as an Application Programming Interface (API) call into the software environment described above.
Model Version(s):
Audio2Face-3D-v3.0
Training, Testing, and Evaluation Datasets:
Training Dataset:
Data Modality
Audio
3D facial motion
Audio Training Data Size
Less than 10,000 Hours
Data Collection Method by dataset
Human - 3D facial motion data and audio
Labeling Method by dataset
Human - Commercial capture solution and internal labeling
Properties (Quantity, Dataset Descriptions, Sensor(s)): Audio and 3D facial motion from multiple speech sequences
Testing Dataset:
Data Collection Method by dataset:
Human - 3D facial motion data and audio
Labeling Method by dataset:
Human - Commercial capture solution and internal labeling
Properties (Quantity, Dataset Descriptions, Sensor(s)): Audio and 3D facial motion from multiple speech sequences
Evaluation Dataset:
Data Collection Method by dataset:
Human - 3D facial motion data and audio
Labeling Method by dataset:
Human - Commercial capture solution and internal labeling
Properties (Quantity, Dataset Descriptions, Sensor(s)): Audio and 3D facial motion from multiple speech sequences
Inference:
Acceleration Engine: TensorRT Test Hardware:
T4, T10, A10, A40, L4, L40S, A100
RTX 6000ADA, A6000, Pro 6000 Blackwell
RTX 3080, 3090, 4080, 4090, 5090
Ethical Considerations:
NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.
For more detailed information on ethical considerations for this model, please see the Model Card++ Bias, Explainability, Safety & Security, and Privacy Subcards.
Please report model quality, risk, security vulnerabilities or NVIDIA AI Concerns here
Bias
Field
Response
Participation considerations from adversely impacted groups protected classes in model design and testing:
None
Measures taken to mitigate against unwanted bias:
None
Explainability
Field
Response
Intended Task/Domain:
Customer Service, Media & Entertainment
Model Type:
Transformer, Diffusion
Intended Users:
Interactive avatar developers, Digital content creators
Output:
Facial pose
Describe how the model works:
Audio input is encoded and concatenated with emotion label, then passed into diffusion-mechanism to output facial motion sequence.
Name the adversely impacted groups this has been tested to deliver comparable outcomes regardless of:
Not Applicable
Technical Limitations & Mitigation:
This model may not work well with poor audio input.
Verified to have met prescribed NVIDIA quality standards:
Yes
Performance Metrics:
Lipsync accuracy, Latency, Throughput
Potential Known Risks:
This model may generate inaccurate lip poses given low-quality audio input.
The Principle of least privilege (PoLP) is applied limiting access for dataset generation and model development. Restrictions enforce dataset access during training, and dataset license constraints adhered to.
Citation
@misc{nvidia2025audio2face3d,
title={Audio2Face-3D: Audio-driven Realistic Facial Animation For Digital Avatars},
author={Chaeyeon Chung and Ilya Fedorov and Michael Huang and Aleksey Karmanov and Dmitry Korobchenko and Roger Ribera and Yeongho Seol},
year={2025},
eprint={2508.16401},
archivePrefix={arXiv},
primaryClass={cs.GR},
url={https://arxiv.org/abs/2508.16401},
note={Authors listed in alphabetical order}
}
