The performance decreases as the length of video increases

[madmaverickminion]

The performance of FSGAN exhibits a decrease when processing videos longer than one minute. Specifically, I have conducted tests using 1080p videos that are around three minutes in length, with minimal variation in head movement and background changes. My objective is to replace a photo with the face of a person in the video.

Upon using the complete video as input, I have noticed that only a portion of the video is successfully modified, up to approximately 1 minute and 15 seconds, beyond which the process fails. However, when I divide the video into two parts near the middle, the FSGAN algorithm functions properly for both segments. It successfully swaps the face with significant frames within each part of the video.

What could be the reason for this behaviour? Is there a limit to the length of the video that we can process using this implementation? @YuvalNirkin @SajjadAemmi