Grain for plastic skin

Give it a take that is right in every way except that skin and surfaces look too smooth (a typical note on H3 Max and other fast models); get back the same take with moving, brightness-neutral film grain, at a strength picked from a ladder. It is the cheapest fix for a texture note: seconds of ffmpeg instead of a re-render. It won't fix shine that is in the lighting, or detail the render never had (hair strands, stubble): for those, re-render on a model with photographic texture.

Inputs

The take at its delivery size: after topaz-upscale if it needs one, since grain size is in output pixels.
Its width, height and frame rate (ffprobe-inspect-media).
The source shot it has to match, and a box on a face or a flat patch (a wall) for the crop sheet, as w:h:x:y at 1080p.

Run it

One command per rung. The noise is generated at half resolution (size 2) and scaled up, so each grain is 2 px wide. For a 1920×1080, 24 fps take at strength 10:

exec ffmpeg -hide_banner -loglevel error -y
  -i <job>/rev1/shot03.mp4
  -f lavfi -i "color=c=0x808080:s=960x540:r=24,format=yuv420p,noise=c0s=10:c0f=t+u,scale=1920:1080:flags=bicubic"
  -filter_complex "[0:v]format=yuv420p[a];[a][1:v]blend=all_mode=grainmerge:shortest=1,format=yuv420p"
  -c:v libx264 -crf 14 -preset slow -pix_fmt yuv420p -c:a aac -b:a 192k
  -movflags +faststart
  <job>/shot03_grain_ladder/shot03_size2_s10.mp4

The noise source is mid grey (0x808080), so grainmerge adds texture without moving brightness.
s= is the take's size divided by the grain size (W/2 × H/2), r= is the take's frame rate, and scale= brings it back to W × H.
c0s= is the strength; c0f=t+u makes it temporal (new every frame) and uniform.
shortest=1 ends the output with the take, so the frame count is unchanged.

Run the ladder at strengths 7, 10 and 14 (name the files <shot>_size2_s<strength>.mp4), then a 100% crop sheet of one frame (frame 12 here) across the source, the take without grain and every rung, left to right:

exec ffmpeg -v error -y
  -i <source shot> -i <take> -i <…_s7.mp4> -i <…_s10.mp4> -i <…_s14.mp4>
  -filter_complex "[0]select=eq(n\,12),crop=384:320:760:60[c0];[1]select=eq(n\,12),crop=384:320:760:60[c1];[2]select=eq(n\,12),crop=384:320:760:60[c2];[3]select=eq(n\,12),crop=384:320:760:60[c3];[4]select=eq(n\,12),crop=384:320:760:60[c4];[c0][c1][c2][c3][c4]hstack=5"
  -frames:v 1 <job>/shot03_grain_ladder/shot03_compare_source_none_7_10_14.png

Look at the sheet, then put the sheet and the rungs on a review page next to the source and let the person pick.

What you get

One grained take per rung and a crop sheet. On the first job, strength 10 at size 2 was picked ("light") for a manager's face that the client called plastic and shiny; 7 is lighter, 14 heavier. Keep the ungrained take beside the picked one.

Tips and traps

Size 2, not 1. Grain one pixel wide vanishes whenever the clip plays scaled down (a review grid, a phone), so the note comes back. Half-resolution noise survives.
Judge at 100%. A review page shows clips scaled down, where every rung looks alike; the crop sheet is the evidence. Open the lightbox at full size for the moving clip.
Grain does nothing for shine baked into the lighting. If the note is about shine, say so before spending time on a ladder.

Check it

The frame count equals the take's (exec ffmpeg -v error -i <out> -map 0:v -f framemd5 -, counting lines that don't start with #).
Measure the same flat patch or face on the source, the take and the pick with python-measure-detail, and compare the numbers against the source's.

Hypernatural