You raise valid technical points that deserve a direct response.
You're right - applying blur to YOLO-detected objects is indeed standard functionality that's been available in libraries for years. Nothing groundbreaking there.
Your concern about blur reversibility is also legitimate. Deblurring algorithms, especially deep learning-based ones, are freely available on GitHub. While many focus on text recovery, facial deblurring is increasingly feasible. Recent papers show promising results in recovering faces from various blur types.
Face replacement or inpainting would provide stronger privacy protection since they destroy the original biometric data rather than just obscuring it. With blurring, the underlying information remains partially intact and potentially recoverable.
However, the value here might be in accessibility rather than innovation. Most people need simple, immediate privacy tools they can actually use without technical expertise. Even imperfect protection beats no protection for everyday users sharing photos online.
For true anonymization, you're correct - replacement or removal would be superior. Generative AI could swap faces while maintaining natural appearance, or inpainting could fill the space with background. These would be irreversible, unlike blur.
The project could acknowledge these limitations upfront and perhaps offer blur as the "quick and easy" option while recommending stronger methods for sensitive use cases.