Faces You Know, Privacy You Keep: Facial Recognition Done Right
Type a name into Google Photos and the app finds every photo of that person across thousands of images in seconds. It is impressive - and most people never stop to think about how that works. The answer is that your face is analyzed, mapped, and compared against a model Google has refined for years using the photos of billions of users. It works so well because you, and everyone you know, have quietly helped train it.
With PixelUnion, you get the same capability. You can search your library by face, group photos by person, and find every picture of your grandmother from 2008 to last Christmas - with accuracy that rivals Big Tech. The difference is in how it works.
Open Source, Not Your Data
Facial recognition in PixelUnion is powered by Immich, the open-source photo platform our service runs on. Immich uses InsightFace, a well-established open-source face analysis library, together with the buffalo_l model for detection and recognition. These are pretrained models developed by researchers and released publicly. The code is public, the models are public, and anyone can inspect how they work.
What that means in practice: when PixelUnion analyzes your photos for faces, it runs these established models on our servers - European servers, under European law. The analysis happens entirely inside your own library. No image, no face embedding, no biometric data ever leaves your PixelUnion environment to be processed by third parties. And crucially: your photos are never used to train or improve models. Not ours, not anyone else’s.
That is a fundamentally different architecture from Big Tech. Google Photos improves its face recognition by learning from users. The more you use it and confirm suggestions, the better it gets - not just for you, but for Google’s model as a whole. You are both the customer and the training dataset.
How Good Can Open Source Really Be?
Skepticism is understandable. When people hear “open-source AI model,” they sometimes imagine something clunky and unreliable - a second-rate substitute. The reality, especially in recent years, is very different.
The InsightFace buffalo_l model behind Immich’s facial recognition is not a toy. It is a state-of-the-art model that performs strongly on standard face detection and verification benchmarks. In independent evaluations, it reaches accuracy comparable to - and in some cases higher than - models behind commercial products.
In daily use, this is easy to see. The system:
- Reliably detects faces even in partial profile, low light, or older low-resolution photos
- Groups the same person across decades of photos while handling changes in hairstyle, weight, and age
- Distinguishes between family members who look very similar
- Performs consistently across different skin tones and lighting conditions
Users regularly describe the results as remarkable. You assign a name, merge a few face groups it has identified, and within minutes years of photos are organized in a way that would have taken hours by hand. This is not a consolation prize compared to Google Photos - it is simply excellent.
Biometric Data Is Sensitive Data
Under the GDPR, biometric data used for uniquely identifying a person is classified as special categories of personal data - the highest protection level in European data protection law. That includes the face embeddings generated by face recognition systems.
Google Photos, operating under U.S. law with global infrastructure, has a different relationship to your biometric data than a European service does. The U.S. has no federal biometric privacy law equivalent to GDPR, although some states have introduced their own rules. Regardless of legal framework, the commercial incentive is clear: face data improves Google’s products, and Google’s products make Google money.
PixelUnion has no such incentive. We do not sell ads. We do not train models. We charge a simple subscription fee for a service that works. Your face data stays in your library, on our servers, under your control.
Privacy and Performance Are Not a Trade-Off
There is a persistent myth that you must choose between a good product and a private one. That the best features only come from companies willing to harvest your data to build them. Facial recognition in PixelUnion is a direct counterexample.
Open-source research has produced models accurate enough to power a genuinely useful feature. That feature runs on European infrastructure, processes your data in isolation, and never uses what it learns about your face to benefit anyone else. No compromise required.
If you have stayed with a Big Tech photo service because you assumed it was the only way to get good facial recognition, it is worth rethinking that assumption.
At PixelUnion, your memories are yours. Every photo, every face, every album - stored on European servers, analyzed on our infrastructure, and never shared with anyone. That is how it should work.