gblur ignores resized input

[Jozef Chutka]

When the input of gblur filter is resized in time, the gblur would keep blurring just the original area (size).

In this example input is resized from 100x100 to 200x200 within 5 frames, however only 100x100 area is blurred in every frame.

Also tried with eval=frame but gblur doesnt seem to have the option.

How to reproduce:

% ffmpeg -f lavfi -i color=size=200x200:rate=5:d=1 -i lines.png -filter_complex "[1]setpts=N/5/TB,fps=5,tpad=stop_mode=clone:stop_duration=1,scale=iw+n*20:-1:eval=frame,gblur=10,[0]overlay" -y out%%d.png

Tested with ffmpeg version 2022-02-07-git-04cc7a5548-full_build-www.gyan.dev

[Michael Koch]

That's a known limitation. Most filters don't accept an input stream where the size changes dynamically. It works only in a few cases, for example scale directly followed by overlay.

Michael

[Michael Koch]

Possible workaround (I didn't test it):
Place your 100x100 object in the center of a 200x200 image with transparent background. Use the zoompan filter for dynamic zooming in, then comes the blur filter (with constant input size), and then overlay.

Michael

[Michael Koch]

Suggestions for improvement:

1. There should be a warning or error message in those cases where dynamic size changes aren't allowed.
2. Add this note to the documentation of the scale filter:

"Please note that although it's possible to change the size dynamically, this doesn't necessarily mean that the subsequent filter accepts a dynamically changing size. This works only in a few cases, for example scale directly followed by overlay."

Does anybody have a "positive" list of cases where dynamic size changes are working? So far, I only know that scale,overlay does work.

Michael

[Jozef Chutka]

Hi Michael,

Thanks for the proposed workaround. I can see eq working ok with dynamic size, however fillborders or colorchannelmixer do not.

With the proposed workaround gblur might blend transparent pixels (from overlay bg) into the main image and fillborders will be problematic as well.

I would like to propose for gblur the following:

• expressions for sigma and sigmaV with available n/t/iw/ih... variables (sendcmd works now but undocumented)
• implement eval=frame so each frame is re-evaluated for expressions + dynamic input size