Getting the frame right for where it is going
A 16:9 video posted to a 9:16 feed is a small letterboxed rectangle in a sea of black. Fixing that is a resize, a crop, or a reframe, and they are not the same thing.
Three numbers describe the shape of a video or an image: its aspect ratio, which is the proportion of width to height; its resolution, which is the pixel count; and its orientation, which is really just which of the first two is larger. Platforms are opinionated about all three, and getting them wrong does not produce an error. It produces a post that looks smaller and worse than the ones around it.
What each place wants
| Destination | Ratio | Resolution | Notes |
|---|---|---|---|
| YouTube, a web page, a TV | 16:9 | 1920x1080, or 3840x2160 | Landscape. The default everything records in. |
| TikTok, Reels, Shorts | 9:16 | 1080x1920 | Full-screen vertical. Anything else is letterboxed. |
| Instagram feed | 4:5 (or 1:1) | 1080x1350 | Taller than square, shorter than a Short. Takes the most screen in the feed. |
| X, LinkedIn, Facebook feed | 16:9 or 1:1 | 1280x720 or 1080x1080 | Square gets more vertical space in a feed than landscape. |
| Stories (any) | 9:16 | 1080x1920 | Keep the top and bottom 250 pixels or so clear of anything that matters. |
Two things are true about this table. The numbers change a little every year. And they all reduce to three ratios: 16:9, 9:16 and something near square. If you can produce those three, you can post anywhere.
Resize, crop, or reframe
These three words get used interchangeably and describe three different operations.
Resizing changes the pixel count while keeping the shape. 1920x1080 to 1280x720 is a resize. Nothing is lost from the edges; detail is lost everywhere evenly. It is what you do when the file is too large or the destination cannot show the extra pixels. The resize tool does this and keeps the aspect ratio by default, because a resize that does not is a stretch, and stretched faces are the first thing anyone notices.
Cropping changes the shape by cutting away the edges. Taking a 1920x1080 frame and keeping the middle 608x1080 of it makes a 9:16 vertical video. Nothing is scaled; you simply lose the sides. This works when the subject sits in the middle of the frame and stays there, and fails when it does not: an interview shot with two people at the sides of the frame, cropped to vertical, shows a wall. The crop tool lets you choose which part of the frame to keep.
Reframing is cropping that moves. The crop window follows the subject across the frame, so a speaker who walks from left to right stays centred in the vertical output. That requires knowing where the subject is in every frame, which is what the editor lets you do by hand, and what the make-a-short tool approximates with a fixed frame and a blurred background.
The blurred-background compromise
When a landscape shot has to become vertical and the crop would lose something important, the common answer is to place the whole landscape frame across the middle of the vertical one and fill the space above and below with a blurred, enlarged copy of itself. It is a compromise. The actual footage occupies a third of the screen. But it loses nothing from the sides, it is instantly recognisable as intentional rather than as a mistake, and it leaves the top and bottom of the frame as soft, low-contrast areas that captions sit on well. For a clip of a wide shot, a game, a screen recording or anything where the edges matter, it is the right call. For a single person talking to camera, crop instead; they were in the middle anyway.
Even numbers, and why the command says -2
If you look at the command a resize produces, the height or width is often written as -2 rather than a number: scale=-2:720. That tells FFmpeg to calculate the missing dimension from the aspect ratio and round it to an even number. The rounding matters because h.264 in the ordinary colour format stores colour at half resolution, so it needs both dimensions to be divisible by two, and an odd width makes the encoder refuse the job. It is the kind of detail that produces a baffling error message when you do it by hand, and it is why the tool writes the command for you.
Upscaling is not resolution
Making a 720p video into a 1080p one by resizing it does not add detail. It spreads the existing detail across more pixels. Every pixel in the result is an interpolation of the pixels around it in the source, and the only thing that changes is the file size, which goes up. Platforms will sometimes treat a 1080p upload better than a 720p one, compressing it less aggressively, and that is a legitimate reason to do it. But the picture will not be sharper, and a viewer comparing it to real 1080p footage will see the difference.
Image upscaling with a model is a different thing. The upscaler here runs a neural network that has learned what edges, textures and text look like at higher resolution and invents plausible detail to match. On photographs and illustrations the result is often genuinely better than interpolation. On a face it is guessing, and it can guess wrong in ways that look fine until you compare with the real person. It is an improvement for a product photo that needs to be larger; it is not a way to recover a licence plate.
Scaling filters
When an image or video is resized, something has to decide what colour each new pixel is. The choices that matter are bicubic, which is the default in most software and is smooth and safe; Lanczos, which is sharper and the better choice when reducing size, at the cost of slight ringing on hard edges; and nearest-neighbour, which keeps every pixel hard-edged and is only right for pixel art. The GIF and image-sequence tools here use Lanczos when shrinking frames, which is where the choice is visible; the plain image resizer uses FFmpeg's bicubic default, which is the safer choice when the output may be enlarged as well as reduced. If a large photo brought down to 1200 pixels looks a little soft, a touch of sharpening afterwards is the usual fix.
Orientation
A phone held sideways records a landscape frame and writes a tag saying "rotate this 90 degrees". Most players honour it. Some, particularly older software and some web players, do not, and the video plays on its side. The rotate tool fixes this the durable way, by actually rotating the pixels and clearing the tag, so the result is the right way up everywhere. It requires re-encoding, since the frames themselves change; a tag-only rotation would be instant but would leave the problem for the next player that ignores it.
A working order
- Decide the destination first. That gives you a ratio.
- If the source is a different ratio, decide whether the edges matter. Crop if they do not; blurred background or a hand reframe if they do.
- Then resize to the destination's resolution, never larger than the source.
- Then compress, last, once. Every step before this one should be done from the best copy you have.