The Animation Motion Dataset is one table of anime shots: one row per shot, with its start, its end and a pixel flux motion score. One Punch Man produced 14,629 rows across 36 episodes, and the same process has since run over Frieren.
PySceneDetect cuts at threshold 27, then each shot gets one pixel flux number
Content detection finds the cuts, then each shot gets the frame-to-frame difference averaged across the shot.
- Below 27, an impact flash counts as two cuts. Above it, quick cuts between similar compositions are lost.
- Anime is harder on that dial than live action, because flashes and colour inversions are ordinary storytelling.
- The three One Punch Man seasons came out around 5.6, 6.0 and 4.9 median motion.
- Season 3 is spread far wider than the other two.
Pixel flux misses camera moves, cheap movement and held frames
The number describes frames, and three things people mean by animation fall outside it.
- A slow pan over a painted background scores high with nothing animated. Raw difference cannot separate a camera move from a character move.
- A held drawing under a shaking effect layer scores well, and smeared in-betweens beat the clean drawings either side.
- A held frame scores the same whether it was a choice or a saving.
- These three explain why the One Frame Man measurement failed and why pixel flux cannot rank Frieren against One Punch Man.
Pixel flux compares episodes fairly within one show
Every episode of one show is measured the same wrong way, so the comparison survives the flaws.
- Across two shows edited on opposite rhythms, it stops working.
- The shots share a clock with the chat data in the One Punch Man dataset.
- That clock lines up audience reaction against animation, second by second.