`--vf=scale` isn't supported after ffmpeg 95568c4 about mpv

[kasper93]

@haasn

[sfan5]

it's FFmpeg/FFmpeg@bb80445
mpv rejects filters with more than one in-/output for use with --vf:

mpv/filters/f_lavfi.c

Lines 980 to 997 in a26bbbd

	// Does it have exactly one video input and one video output?
	static bool is_usable(const AVFilter *filter, int media_type)
	{
	#if LIBAVFILTER_VERSION_INT >= AV_VERSION_INT(8, 3, 0)
	int nb_inputs = avfilter_filter_pad_count(filter, 0),
	nb_outputs = avfilter_filter_pad_count(filter, 1);
	#else
	int nb_inputs = avfilter_pad_count(filter->inputs),
	nb_outputs = avfilter_pad_count(filter->outputs);
	#endif
	bool input_ok = filter->flags & AVFILTER_FLAG_DYNAMIC_INPUTS;
	bool output_ok = filter->flags & AVFILTER_FLAG_DYNAMIC_OUTPUTS;
	if (nb_inputs == 1)
	input_ok = avfilter_pad_get_type(filter->inputs, 0) == media_type;
	if (nb_outputs == 1)
	output_ok = avfilter_pad_get_type(filter->outputs, 0) == media_type;
	return input_ok && output_ok;
	}

[haasn]

Does this fix it?

commit ce996d2d80a6d184856c9d3a869a10e8f1414e9b
Author: Niklas Haas <[email protected]>
Date:   Fri May 3 22:07:30 2024 +0200

    avfilter/vf_scale: add missing filter flag
    
    Fixes: bb8044581366fe286e16b14515d873979133dbda

diff --git a/libavfilter/vf_scale.c b/libavfilter/vf_scale.c
index bc53571c1c..696ee272ec 100644
--- a/libavfilter/vf_scale.c
+++ b/libavfilter/vf_scale.c
@@ -1261,6 +1261,7 @@ const AVFilter ff_vf_scale = {
     FILTER_QUERY_FUNC(query_formats),
     .activate        = activate,
     .process_command = process_command,
+    .flags           = AVFILTER_FLAG_DYNAMIC_INPUTS,
 };
 
 static const AVFilterPad avfilter_vf_scale2ref_inputs[] = {

[sfan5]

I bet it does but isn't that misusing the flag?

/**
 * The number of the filter inputs is not determined just by AVFilter.inputs.
 * The filter might add additional inputs during initialization depending on the
 * options supplied to it.
 */

I suppose scale could be changed to dynamically add the ref input if any of the relevant vars are used.

Edit: Based on the ffmpeg documentation I noticed we're not actually checking the pad count correctly and proposed #14052. Incidentally this would make your patch no longer work.

[haasn]

The pad count serves as a minimum even if the filter has dynamic inputs/outputs, so we should reject them here.

Huh, then indeed vf_scale would need to be fixed to only add the second input dynamically if it's being used. I'm not sure if I read the same into that documentation as you, though. You think that implies it can only ever strictly add inputs, rather than also removing them?

[sfan5]

Yes, that is my interpretation.

[zhongfly]

Fixed by FFmpeg/FFmpeg@6a5b021