Practical
Upscaling and Downscaling: What Resizing Really Costs
Making a video smaller is nearly free and often improves it. Making it larger adds no detail and usually makes it worse. Here is what actually happens either way, and when each is worth doing.
3 min read · Published 25 August 2026
Resizing a video sounds symmetrical: bigger or smaller, one setting either way. It is not. The two directions are completely different operations with completely different outcomes.
Downscaling: throwing away information you have
Going from 4K to 1080p means every four pixels become one. The software has real data for all four and produces a considered average.
The result is often better than footage shot at 1080p to begin with. Averaging four samples reduces noise, sharpens apparent detail, and hides small focus errors. This is why professional workflows routinely shoot larger than they deliver.
It also compresses far better. Noise is the most expensive thing to encode, and downscaling removes a great deal of it, so a downscaled 1080p file at a given bitrate looks cleaner than a native one.
Upscaling: inventing information you do not have
Going from 1080p to 4K means one pixel becomes four, and the software has to decide what the three new ones contain. There is no source for that decision. It is guessing from the neighbours.
Traditional upscaling produces a larger, softer version of the same picture. The detail was never captured and no arrangement of arithmetic recovers it. What it definitely does is quadruple the pixel count, which makes the file substantially larger for no gain.
| Downscaling | Upscaling | |
|---|---|---|
| Detail | Preserved, often improved | None added |
| Noise | Reduced | Magnified |
| File size | Much smaller | Much larger |
| Compresses | Better | Worse |
| Reversible | No | No |
What about AI upscaling?
Modern upscalers trained on large amounts of video do produce genuinely better results than simple interpolation, and on some material they are impressive.
It is worth being precise about what they do: they generate plausible detail, they do not recover real detail. On faces and text a good model produces something convincing. On anything unusual it can produce confident nonsense, and it is not always obvious which you are looking at.
For personal footage that matters more than accuracy, that trade can be fine. For anything documentary, evidential or archival, generated detail is a problem rather than a feature.
Getting downscaling right
- Scale to standard resolutions. 1920 by 1080 and 1280 by 720 are handled well everywhere; odd sizes cause trouble with some encoders and players.
- Keep the aspect ratio. Scaling one dimension without the other distorts everything, and it is very obvious on faces.
- Keep dimensions even. Most codecs require even numbers and will silently round, which shifts the picture by a pixel.
- Downscale once, from the largest source you have. Two smaller steps lose more than one direct one.
- Reduce the bitrate to match. A quarter of the pixels does not need the original bitrate, and leaving it high wastes a great deal of space.
Scale before compressing, not after. Downscaling a compressed file magnifies its artefacts, since the encoder's blocks and banding get treated as real picture content by the scaler.
When upscaling is actually the right call
There are a few genuine cases. A platform that allocates a more generous bitrate to higher tiers. A destination that requires a specific resolution and will do a worse job of the conversion itself. Mixing sources in one timeline where everything must share a resolution.
In every one of those, the reason is compatibility or delivery, not quality. Upscaling for quality is the one reason that never works.
The rule worth keeping
Capture at the highest resolution you reasonably can, keep that as your master, and produce smaller versions from it as needed. Every version you make from a large original is good. Every version you make from a small one is limited by it, permanently.
Common questions
Does upscaling improve video quality?
No. The detail was never captured, so scaling up can only guess from neighbouring pixels. You get a larger, usually softer file. AI upscalers generate plausible detail, which is a different thing from recovering real detail.
Is downscaled 4K better than native 1080p?
Usually yes. Averaging four pixels into one reduces noise and sharpens apparent detail, and the result compresses better too, which is why shooting larger than you deliver is standard practice.
Should I resize before or after compressing?
Before. Downscaling an already compressed file magnifies its artefacts, because the scaler treats encoder blocks and banding as real picture content.
Why do odd dimensions cause problems?
Most codecs work in blocks and require even numbers, so an odd dimension gets silently rounded. That shifts the picture by a pixel and can produce a thin band at an edge.