The Still-Frame Threshold Depends on Bit Depth

The short version

The YAVG threshold scales with the material’s bit depth.

Our measured values were these.

Export	pix_fmt	YAVG counted as still
ProRes 4444	yuv422p12le (12-bit)	4.8
ProRes 422 HQ	10-bit	1.2

Exactly a factor of four apart. 12-bit luma runs 0–4095 and 10-bit runs 0–1023 — the range itself is about four times larger. For the same difference in the picture, a different bit depth changes the YAVG number by that same factor of four.

So applying 4.8, meant for 12-bit, to 10-bit material classifies moving frames as still. And applying 1.2, meant for 10-bit, to 12-bit makes still material read as moving.

What is actually happening

Detecting still frames, done plainly, looks like this.

1. Build the difference between consecutive frames (tblend’s difference)
2. Read the mean brightness of that difference image (signalstats’ YAVG)
3. Treat stretches where the mean is below a threshold as “still”

The less difference there is, the blacker the difference image and the closer YAVG gets to 0. “How black is it” is standing in for “is it still”.

So YAVG is a luma value itself, not a ratio. Change the range and the number changes.

Check before you use it

Before applying a threshold, look at how many bits the material is.

ffprobe -v error -select_streams v:0 \
  -show_entries stream=pix_fmt,codec_name \
  -of default=nw=1 input.mov

codec_name=prores
pix_fmt=yuv422p12le

12le is 12-bit, 10le is 10-bit.

It changes even when you did not mean to change the export settings. Switching the codec preset from ProRes 4444 to 422 HQ alone drops the bit depth, so measuring with the threshold left as it was quietly breaks the verdict from that day onward.

No error appears. The numbers keep coming out. They keep coming out while meaning something different — that is what makes this kind of measurement dangerous.

Decide the threshold by measuring

Rather than reusing an inherited value, take the boundary on your own material.

1. Pick a clearly still stretch and measure its YAVG
2. Pick a clearly moving stretch and measure its YAVG
3. Put the threshold between them

If the two values are too close, that material cannot be separated by YAVG. Rather than persisting with threshold tweaks, admitting it cannot be separated finishes sooner.

Split long videos

Push a long piece through in one go and you will not notice when the render dies partway.

Ours is four parts of 1080 frames each, each run in the foreground before being concatenated. The urge to send it to the background is strong, but moving on to the next step without noticing it died costs more.

Do not confuse numbers coming out with measuring

The most dangerous thing about this measurement is that the results come out perfectly normally even when the threshold is wrong.

An output of “0 stills” does not distinguish between there being no stills and the threshold being too loose to catch any. The same screen appears either way.

Checking pix_fmt every time is how a human keeps hold of that distinction.