The domain of images is massive. Think about the number of images you can create if you set each pixel of a 32x32 image to a value from 0-255. (The number of images you can create is “infinite.”) Because of this there are corner cases that may have not been covered by the models’ training or the image has been manipulated in such a way that it conceals its signature features that the detectors are looking for.
We are always improving our models. It is like cybersecurity where malicious hackers try to come up with new and novel viruses to corrupt systems, while cybersecurity companies catch up and develop antivirus software. The world of deepfakes is similar. People come up with new and novel ways of generating deepfakes and Reality Defender tries to preempt this by thinking ahead of the curve, being ahead of the trends and reacting quickly to new generative methods.