Short answer: to pad a vertical video into a 1920×1080 frame with FFmpeg, scale it so the longer side fits, then add bars with pad. A bare scale=1920:1080 stretches the picture. A 16:9 crop fills the frame but throws away the top and bottom. The five pytest checks below were run on a real 12.064-second 720×1280 fixture (grey-hoodie-v2.mp4, md5 a50e13313b71a76dbd179dcdc2c2b6bf).

This is a reproducible FFmpeg/Python check, not a product walkthrough. An optional later CTA points at Voor's Gangnam Style AI Dance page if you want to inspect a current image-to-video form after you have a 16:9 file.

Why scale=1920:1080 is the wrong first command

The fixture is 720×1280, a 9:16 portrait. A 1920×1080 canvas is 16:9. Those aspect ratios do not match, so there are three different operations that all happen to emit a 1920×1080 file:

  • Pad: scale the source until it fits inside 1920×1080, then fill the leftover columns (or rows) with a solid color, usually black. The people and text in the frame keep their shape.
  • Stretch: scale=1920:1080 with no aspect flag maps every source pixel onto the new rectangle. The same person becomes wider.
  • Crop: cut a 16:9 window out of the portrait, then scale that window up. The output is filled edge to edge, but the headroom and the lower body leave the frame.

On this fixture the fitted size is 608×1080 (720/1280 × 1080 = 607.5, rounded to an even width for yuv420p). The pad then places that 608-pixel-wide picture at x=656, with black columns on both sides.

A 720 by 1280 source frame of a person in a grey hoodie, next to the same frame scaled to fit and padded to 1920 by 1080 with black bars on both sides, preserving the original proportions

Scale to fit, then pad

The filter chain that the tests run is:

ffmpeg -i fixture.mp4 -vf "scale=1920:1080:force_original_aspect_ratio=decrease,pad=1920:1080:(ow-iw)/2:(oh-ih)/2" -c:v libx264 -pix_fmt yuv420p -c:a copy padded.mp4 

force_original_aspect_ratio=decrease is the part that prevents stretching. pad=1920:1080:(ow-iw)/2:(oh-ih)/2 then centers the fitted picture. -c:a copy leaves the AAC stream untouched; this is already a re-encode of the video, so you cannot add -c:v copy on the same command.

After the command, ffprobe reports 1920×1080, h264 plus aac, and a duration within 0.05 seconds of the source 12.064 seconds. The AAC stream md5 matches the source. The left and right bars on the first frame are black, and the 608×1080 center crop stays close to a scale-only first frame of the same source (mean absolute error under 8 after the yuv420p encode).

Stretching and cropping are different 1920×1080 files

To show the two near-misses, the tests also write a one-frame stretch and a one-frame crop:

ffmpeg -i fixture.mp4 -vf "scale=1920:1080" -frames:v 1 stretched.mp4 ffmpeg -i fixture.mp4 -vf "crop=720:405:(iw-720)/2:(ih-405)/2,scale=1920:1080" -frames:v 1 cropped.mp4 

Both outputs are 1920×1080. Neither has black side bars. In the stretched file the subject is visibly wider than in the source. In the cropped file a 720×405 window (16:9) was taken from the middle of the 720×1280 frame, so the top and bottom of the original picture are gone. If you wanted the full body in a 16:9 canvas, crop is the wrong tool.

The same 720 by 1280 source frame stretched to 1920 by 1080 with no aspect-ratio flag; the person appears wider than in the original

What the tests check

The pytest file in this post runs five checks against the fixture and the commands above:

  1. Padding with scale=...force_original_aspect_ratio=decrease,pad=1920:1080:(ow-iw)/2:(oh-ih)/2 produces a 1920×1080 H.264/AAC MP4 whose duration stays within 0.05 seconds of 12.064 seconds.
  2. -c:a copy keeps the AAC stream md5 identical to the source.
  3. The left and right bars on the first padded frame are black, the 608×1080 center is not, and that center stays close to a scale-only fitted frame of the same source.
  4. A bare scale=1920:1080 also writes 1920×1080, but its left edge is content rather than a black bar, and the frame does not match the padded file.
  5. Cropping to 720×405 and then scaling to 1920×1080 also fills the canvas with no side bars, so the top and bottom of the portrait have been discarded.

Run them from the directory that contains test_pad_video.py and inputs/grey-hoodie-v2.mp4:

python3 -m pytest -q test_pad_video.py 

After you have a 16:9 file

Once the padded MP4 is 1920×1080 and the subject still has the original proportions, you can drop it into an editor or a player that expects a landscape frame. If you then want to look at an image-to-video form that already targets a short clip, Voor's Gangnam Style AI Dance page is a current landing page you can inspect without using the commands above.