Pixelation has been used as a privacy tool in journalism and broadcasting for decades. A face pixelated on television in the 1990s was genuinely unrecognisable to any viewer. In the era of AI image enhancement, the picture is more nuanced.
Research published by computer vision teams has demonstrated that faces pixelated at block sizes of 8-10px can sometimes be partially reconstructed by trained enhancement models, particularly when the original image is high resolution and the face occupies a large area of the frame. The same enhancement fails on faces pixelated at 20px or larger because there is simply not enough spatial information preserved for the model to work with.
The practical implication is that pixelation remains highly effective for the vast majority of real-world use cases, as long as you use a sufficiently large block size. For most sharing contexts where the threat model is a human viewer rather than a dedicated AI adversary, even 10px blocks are more than adequate. For situations where genuine privacy matters, including journalistic sources, vulnerable individuals, and legal or HR documentation, use 20px or larger and consider whether the pixelation is the only privacy measure you need or one layer among several.
Pixelation versus blur: which is more effective?
At equivalent levels of visual obscuring, pixelation and Gaussian blur offer similar practical privacy protection. Pixelation preserves more information about the spatial structure of the content (you can see that something is a face-shaped object even if you cannot identify whose face it is) while blur loses that structure more completely at higher intensities.
For maximum privacy in region-based censorship, combining both techniques is more effective than either alone. Apply blur first, then pixelate on top, which removes the smooth Gaussian structure that enhancement models use as a starting point. Both tools are available separately on this site if you want to use them in sequence.
