Sleap labels offset from display frame by ~20 frames for certain videos

[tbashaw712]

I am using SLEAP to estimate the position of finches in a behavioral task. i have 14 individuals, each with roughly 3.5 hours of footage across multiple videos. for a couple of these videos, I have had issues where SLEAP seems to be using my trained model well to estimate the pose of the animal, but the skeletons that are placed do not match up visually with the frame that is displayed in the GUI. However, they do seem to correlate accurately with the bird's movement, just at a slight time delay to what frame is being displayed within the GUI. This is not the case for every video in my dataset, only some of them. What is causing this, will it be a problem in analysis, and how can I fix it for the affected videos? included are screenshots of the problem in one video in tracking a bird changing position, and another example from an unaffected video of the accurate skeleton tracking.

example of correct tracking lining up with display frames. bird adjusts position and skeleton updates accordingly.

example of correct tracking that is desynced from the display frames within the GUI. the skeleton labeling the bird moves before the bird does by about 20-30 frames.

[tbashaw712]

To add more information - these videos have been re-encoded using the command line ffmpeg -y -i "input.mp4" -c:v libx264 -pix_fmt yuv420p -preset superfast -crf 23 "output.mp4" and the problem persists.