How to upscale a video without it distorting people's faces

Last updated 2026-09-10 · 8 min read

What is the short version?

An upscaler invents detail. On the person you punched in on that is exactly what you want; on a row of faces twenty feet behind them it invents features those people do not have, and the result is subtly, uncomfortably wrong.

1. Go to Vizard Agent with the footage.
2. Say what you want sharper and why.
3. Say no facial features may be altered.

What do you need before you start?

The footage and a clear boundary. "Make it sharper" is a request with no edges; "sharpen the speaker, do not touch anyone in the background, and do not alter anybody's features" is three boundaries that can each be checked afterwards.

• The footage at its original resolution.
• What you want sharper. The subject, usually.
• What must not change. Faces, especially in the background.
• Whether lightening is allowed. Often it is enough.
• Where it will be seen. A big screen is stricter.

What do you type into Vizard Agent?

Set the limit in the same sentence as the request. Enhancement is a spectrum and the default is somewhere in the middle, which is fine for a wall and not fine for a face — saying so up front is cheaper than undoing it later.

Prompt

Variants worth knowing:

• "Do not change facial features." The boundary that matters.
• "Slight lightening or sharpening." How far you will go.
• "A short test first." Cheap, and settles it.

What does Vizard Agent actually do?

Here is the order Vizard Agent worked in on a real recorded event whose punch-ins needed sharpening. Notice that the upscale is never simply applied — it is tested, compared, and then applied to part of the frame.

1. Checks whether an upscaling tool is available and what it costs.
2. Reads the upscaler's details before wiring it in.
3. Cuts the relevant excerpts into one file and uploads it.
4. Starts the enlargement running in the background.
5. Checks why it failed rather than simply retrying.
6. Tests on a ten-second piece to see whether length was the problem.
7. Compares a tight punch-in from the original against the enlarged one.
8. Puts the two side by side at the same moment.
9. Tests a blend that enlarges the speaker and leaves the rest original.
10. Runs the enlargement on all the excerpts in parallel.
11. Rewires the build to blend the result in, and tests it on the sample.
12. Compares the previous punch-in against the new one at the same frame.
13. Verifies every enlarged excerpt before it goes into the cut.
14. Tests brightness lifts on faces with close-up before and after.

Step nine is the answer to the whole problem. An upscale does not have to apply to the whole frame — keeping it on the subject and leaving the background as filmed gives you the sharpness where you were looking and no invented faces anywhere else.

Step six is worth doing on any expensive process. A ten-second test tells you whether the failure was about length, whether the quality is worth the cost, and what the result actually looks like, for a fraction of the whole job.

What does the result look like?

Punch-ins that hold up at the size they are shown, with the speaker sharper and everybody behind them exactly as the camera recorded them. Nobody's features have been reinterpreted, and the brightness lift stops well short of retouching.

Played next to the original, the difference is detail rather than a different-looking person.

When does this not work well?

Upscaling cannot recover what was never recorded. Very soft footage, a heavy crop into a small area, or a face that is a dozen pixels wide gives the tool nothing to work from, so it invents — and inventing a face is the failure you are trying to avoid.

• Very small faces. There is nothing to enlarge.
• Heavy compression. Blocks get sharpened into detail.
• Fast motion. Enlarged frames flicker against each other.
• A crowd in focus. Every face is at risk, not just one.
• Any use where accuracy is legal. Invented detail is not evidence.

How do you fix a result that came back wrong?

Say whose face changed and where in the frame they are. Vizard Agent keeps the side-by-side comparisons and the blend it built, so pulling the enlargement back off the background is an adjustment to that blend rather than a re-render of everything from scratch.

• "The faces behind him look wrong." The blend limited to the speaker.
• "It looks over-sharpened." Eased, and compared again.
• "Undo the enlargement." Rebuilt from the original footage.
• "It is still soft." A tighter blend, or a different approach.

How does Vizard Agent compare to doing it yourself?

The usual route is to run the upscaler on the whole file and look at the result, which is expensive and treats every part of the frame as equally worth inventing. The background faces are also the last thing anybody checks, because the eye goes to the subject.

Common questions

Will it change faces? Not if you say so. Vizard Agent keeps the enlargement off them.

Can it sharpen only the speaker? Yes. Vizard Agent blends the enlargement onto part of the frame.

Can I see a test first? Yes, and you should. Ten seconds is enough for Vizard Agent to show you.

How do I undo it? Vizard Agent rebuilds from the original footage; nothing is baked in.

Is lightening safer than upscaling? Often, yes, and Vizard Agent tests it on faces close up.

What if the footage is very soft? Vizard Agent will say when there is nothing to recover.

Will it flicker? It can on fast motion. Vizard Agent checks frames in sequence.

Does it cost more than a normal render? Yes. Vizard Agent prices it before starting.

Can it enlarge only some shots? Yes. Name them and Vizard Agent leaves the rest.

What about logos and banners in shot? Vizard Agent can leave those on the original too.

Will the colours change? No, unless you ask Vizard Agent for a grade as well.

Can it do this at 4K? Yes, though Vizard Agent notes the cost and the risks both rise.

Should I upscale before or after cutting? After. Vizard Agent enlarges only the excerpts you actually use.

Why does an upscaler distort faces? Because it fills in detail that was never recorded, and on a face the invented detail is a different person's.