I came across a video with a broken frame rate that had choppy playback. The RIFE interpolation I always use didn’t improve the situation, but Aion worked miracles; the resulting video turned out very smooth, thank you for that.
However, there is an issue when the scene detection feature is enabled: the output video ends up being longer than the original. If you compare them side-by-side, there is a progressively increasing desync of both audio and frames (between the two videos). At the same time, there is no audio desync within the output video itself, but there are extraneous audio artifacts (“clicks” or “pops”).
After opening it in a video editor, I didn’t spot any anomalies in the video itself (other than the longer duration). However, when I broke the file down into individual frames, I discovered that at hard scene cuts, the last frame before the scene change is always duplicated. I believe this is the root of the problem. I think the logic for splitting and merging frames is broken here, resulting in the exact same frame remaining both at the cut point and at the point where it should ultimately be stitched back together. I think this isn’t an issue with the model or the scene detection feature itself, but simply an incorrect assembly of the final video file.
Video comparison (for the test, I ran the interpolation with the scene detection feature enabled and disabled):
This is how it looks in the editor:
The files that were used in the comparison test video:
sceneON.zip (46.9 MB)
sceneOFF.zip (47.1 MB)
The last 3 frames before a hard scene cut and the 3 frames after (frame 117 and 118 identical, and this happens across all scenes with a hard cut):
frames2.zip (4.9 MB)
This issue was already raised in 2025 but hasn’t received a fix yet.
If anyone is interested in how Aion works on a broken frame rate, you can download a sample here (unrelated to the issue).
aion.zip (37.6 MB)
