Measure judder

Give it one or more videos; get back, per video, a judder score, the median motion step and the number of repeated frames. A frame's motion step is the mean absolute change in brightness from the frame before. Steady motion keeps the step even, so the score sits near 0. A cadence that repeats one frame and then jumps two scores high. Compare takes of the same shot, or a take against the source footage of that shot. Don't compare scores across different shots: the number depends on how much moves.

Run it

Write references/judder.md's script into the project folder as judder.py, as is, and run it with exec python:

exec python judder.py --videos <source shot>.mp4 <shot>_kling_a.mp4 <shot>_wan_a.mp4
exec python judder.py --videos <take>.mp4 --frames 78 --steps

• --frames N scores only the first N frames. Use the shot's length, so a still hold at the end of an untrimmed render doesn't count.
• --steps also prints every frame's step, to find where a stutter happens.

What you get

JSON per video: judder, median_step, frames, repeats (frames whose step is under 0.4 of the local median, so they repeat the frame before).

Reading the numbers

• Source footage scored about 0.05, and Kling v3 takes about the same.
• A model whose output was 24 fps content padded to 30 fps (one repeat and one double step in every 5 frames) scored 0.78. After the repeats were dropped and the gap motion-interpolated it scored 0.39, and it still looked jerky. A score several times the source's is visible.
• repeats near a fifth of the frames, at a 30 fps container, is that padding pattern. Check the frame rate (ffprobe-inspect-media) and suspect the provider's output before the motion itself.
• A near-static shot (a talking head; a median step around 0.2) gives a noisy score, or none when the median is 0. Judge those by eye.

Tips and traps

• Measure trimmed takes, or pass --frames: a hold after the action lowers the median and inflates the score.
• A cut inside the clip is one huge step and pushes the score up. Measure one shot at a time.
• Retiming a take (speeding it up to fit) adds judder of its own. Trim it, or re-brief it to act in real time, instead.
• Put the source's score next to each take's on the review page, so the person sees the comparison and not a bare number.

Check it

Run it on the source shot first. It should read near 0.05 with 0 repeats. If it doesn't, the source itself is uneven, and takes have to be read against that number.