Introducing AuraFace: Open-Source Face Recognition and Identity Preservation Models

Today, we're excited to announce the release of a new identity encoder model that is commercially available: AuraFace. This model represents a step forward in making advanced face recognition and identity preservation technologies accessible for commercial use.

AuraFace: Open-Source Face Recognition for Commercial Use

AuraFace is our open-source version of the popular ArcFace model, designed specifically for commercial applications. Unlike the original ArcFace, which has non-commercial restrictions due to its training data, AuraFace can be used in commercial projects without limitations.

However, the significant size difference between the training data means that AuraFace does not match the performance of the original ArcFace, but it offers a strong balance between accuracy and commercial usability.

AuraFace IP-Adapter: Preserving Identity in Image Generation

Complementing AuraFace, we've also developed AuraFace IP-Adapter, for maintaining identity consistency in image generation tasks. AuraFace IP-Adapter ensures that the core identity of a subject is maintained across various image generation and manipulation tasks. It is designed to work seamlessly with SD1.5 and can be easily incorporated into existing workflows.

Performance Comparisons

To provide transparency and help users make informed decisions, we've conducted comparisons between the original ArcFace and our AuraFace by training IP-Adapter on AuraFace and comparing:

From Left to Right: Original, IP-Adapter-AuraFace, IP-Adapter-ArcFace

These comparisons demonstrate:

1. The models do not achieve perfect photorealism and ID consistency.
2. The efficacy of the model in subject preservation may vary on the basis of ethnicity.
3. The generalization of the models is limited due to limitations of the training data, base model, and face recognition model.

Evaluation

AuraFace is based on the resnet100 architecture as the original ArcFace model, hence we can compare it to the original in the following metrics:

Real-World Applications

The commercial-friendly nature of AuraFace opens up a wide range of applications:

• E-commerce and Retail: Implement secure facial recognition for payment systems or personalized shopping experiences.
• Digital Content Creation: Use the IP-Adapter for creating consistent digital avatars or characters in games and interactive media.
• Mobile Applications: Integrate face recognition features into apps for enhanced user experiences and security.
• Corporate Security: Develop employee authentication systems without concerns about licensing restrictions.

Open-Source Community and Future Development

As an open-source project, AuraFace's future development will greatly benefit from community contributions. We encourage developers and researchers to:

• Experiment with the model and share their findings, especially related to ethnicities for us to improve our dataset.
• Contribute to expanding the training dataset while maintaining commercial usability.
• Propose and implement improvements to the model architecture. We currently follow the same architecture and training procedure as ArcFace.

Try It Out

AuraFace is now available on HuggingFace. We've also deployed them to dedicated endpoints at fal.ai/lora for easy integration into your projects.

We're excited to see how the community will leverage these new, commercially-friendly models. As we continue to improve AuraFace, we're actively seeking to expand our face dataset. If you're interested in contributing to this open-source effort or have access to diverse face data that could enhance our model's performance, we'd love to hear from you. Reach out to us on Twitter at @_yatharthg, @Gothos03 or @jfischoff to discuss how you can help shape the future of open-source face recognition technology. Your contributions could play a crucial role in making AuraFace even more robust and versatile for commercial applications worldwide!