Thank you!
I hadn't noticed the issue, so this really helps. I also understand the cause of the problem in #1238 now.

However, I'm not actually sure this is a proper fix as I don't think Unity technically guarantees that graphics operations will be finished within one frame.

I agree. It seems like a good idea to leave a FIXME comment.
To fix it properly, we might need to use GlSyncToken to make MediaPipe wait for the copy operation to complete (I am not confident that it will work, though).
ref. https://github.com/google-ai-edge/mediapipe/blob/cf6e4ec469e04e126cc36c1aebc51b48ffdd80c7/mediapipe/gpu/gl_texture_buffer.h#L129-L132

Done, let me know if this works for you.

Completely unrelated, but I've noticed another thing. Currently, inference runs every frame even if the webcam didn't supply a new frame. If the app FPS is higher than the webcam FPS, this causes the same webcam frame to be sent to MediaPipe multiple times.

You could check .updateCount on the texture you get from imageSource.GetCurrentTexture() and if the count wasn't increased, you just yield return null; continue; and check again in the next app frame.

I have a webcam running at 30 FPS and the test app running at 60 FPS (testing on Windows). Implementing this cuts the CPU usage roughly in half, as you would expect:

Of course it only improves performance if the webcam (or other video source) FPS is lower than the app FPS. Otherwise it makes no difference.

If you'd like, I can open another PR for this later, implemented like this: 2f35f73

Thank you for your suggestion!
You're absolutely right that this check would be valuable in a real-world application. However, since this is just a sample app, I believe we should focus on including only the code necessary to run MediaPipe. Therefore, I think it's fine to leave it as is.