When I started developing DeepFace back in 2019, building a state-of-the-art facial recognition pipeline was surprisingly painful. You had to manually implement complex architectures, hunt down pre-trained weights, and deal with inconsistent APIs. DeepFace was born to solve this—a lightweight wrapper designed to run models like VGG-Face, Facenet, and OpenFace in just a single line of code.
At the time, TensorFlow was the undisputed king of deep learning, so building DeepFace on top of TF and Keras backbones was the natural choice. Over the past five years, that simple wrapper evolved into a cornerstone tool for the AI community, recently crossing 15 Million downloads and 23K GitHub stars. But as the library grew, the underlying framework landscape was undergoing a massive paradigm shift!
🙋♂️ You may consider to enroll my top-rated machine learning course on Udemy
The Great Shift: What Papers with Code Data Tells Us
If you look at the historical data from Papers with Code and top AI conferences (CVPR, NeurIPS, ICCV):
- In 2019–2020: TensorFlow and PyTorch were in a neck-and-neck race. TensorFlow held a massive share of real-world deployments, while PyTorch was just beginning to capture the research community.
- Today: PyTorch powers over 80% of published AI research papers and the overwhelming majority of models on Hugging Face.
This shift divided the ecosystem into two distinct worlds:
- Who uses PyTorch today? Researchers, open-source contributors, Generative AI labs, and developers who prioritize rapid prototyping, dynamic debugging, and cutting-edge vision/NLP architectures.
- Who uses TensorFlow today? Enterprise MLOps teams, legacy production pipelines, mobile/edge deployment engineers (via TFLite), and Google Cloud ecosystem users.
As Deepface grew, keeping it strictly bound to TensorFlow meant locking out a massive chunk of the PyTorch-native research and developer community.
What’s New in DeepFace?
Starting with v0.0.101, DeepFace is officially framework-agnostic. It can run seamlessly on PyTorch as well as TensorFlow.
To support this without bloating your environment, we are introducing explicit backend extras during installation:
# Install Deepface with TensorFlow backend pip install deepface[tensorflow] # Install Deepface with PyTorch backend pip install deepface[pytorch]
So, the default “pip install deepface” command will not install tensorflow dependency anymore. To avoid breaking change, this will continue to pull tensorflow for a while, and once you import deepface, you will be warned about this breaking change for the future release.
Future-Proofing Face Detection: The YOLO Ecosystem
One of the most exciting outcomes of this shift is face detection. While recognition models extract identities, high-accuracy face detection is the critical first step in any pipeline.
In recent years, the YOLO (You Only Look Once) family has evolved at a blistering pace—moving from YOLOv8 to newer iterations like YOLOv11. The vast majority of these state-of-the-art vision detectors are built natively on PyTorch.
By unlocking PyTorch support in DeepFace, we are opening the door to seamlessly integrate modern, ultra-fast YOLO-based face detectors as they get released by the community. This ensures DeepFace stays at the absolute bleeding edge of speed and detection accuracy without forcing cumbersome model conversions.
Why This Matters
By supporting both backends, DeepFace offers the best of both worlds:
- No forced dependencies: You only download the framework you actually use.
- Seamless integration: PyTorch projects no longer need heavy workaround pipelines or forced TensorFlow installs to leverage Deepface’s face recognition capabilities.
- Faster innovation: PyTorch support allows us to easily integrate state-of-the-art vision backbones coming out of recent research papers.
We expect this change to ignite a new wave of growth and community contributions. Try out v0.0.101 today and let us know what you think! And please star the repo because it will help us to reach more people!
Support this blog financially if you do like!
