How to replace a person in a video with someone from a photo

Last updated 2026-08-16 · 6 min read

What is the short version?

This is a short technical job with a large consent question wrapped around it. Technically it is a test frame, a comparison and two passes over the footage. Everything difficult about it is who is in the photo and who is in the video.

1. Go to Vizard Agent and upload the clip and the photograph.
2. Say clearly which person in the clip is being replaced and by whom.
3. Say that the movement, the scene and the audio should stay as they are.

What do you need before you start?

Both files and permission from both people. Vizard Agent handles the technical side by testing before committing, and the part it cannot handle is whether the person in the photograph and the person in the video have agreed to this.

• The clip. Short works best. This is expensive per second and the quality falls off with length.
• A clear photograph. Front-facing, well lit, unobstructed. This sets the ceiling.
• Which person. If more than one appears, say precisely which.
• Permission. From the person being removed and the person being inserted. Both.
• What it is for. A gift, a joke among friends, a piece of fiction. It affects how it should be labelled.

What do you type into Vizard Agent?

Say who replaces whom and what must not change. Vizard Agent looks at both files itself to identify the people involved, so a short instruction naming the swap and asking for everything else to be preserved is all it needs before it starts testing on a single frame.

Prompt

Variants worth knowing:

• Test first. Ask for a single frame or a three-second test before the whole clip; it is much cheaper.
• Just the face. Say if the body and clothing should stay as filmed.
• A short section only. Naming the seconds keeps both the cost and the risk down.

What does Vizard Agent actually do?

Here is the order Vizard Agent worked in on a real replacement. The whole shape of the run is caution: one frame, then two versions of that frame, then a three-second test, and only then the full clip in pieces.

1. Inspects the video and the photograph and looks at both.
2. Extracts frames from the clip and reviews them.
3. Checks the available tools and their prices before committing to anything.
4. Chooses a frame where the face is clearest and previews the candidates.
5. Uploads that reference frame and runs the swap on it in two versions.
6. Compares both against the original and looks at the result.
7. Runs the full clip, reads the error log when it fails, and checks the tool's duration limits.
8. Prepares a three-second test and confirms the approach works at that length.
9. Splits the clip into two segments, processes both, compares each against the original, and rejoins them with the original audio.

Step 5 is the habit worth copying. Two versions of a single frame cost almost nothing and told Vizard Agent whether the likeness was going to hold before any real work was spent on it.

What does the result look like?

From the run this page is written from, probed on the delivered file: 576x1024, H.264, 30fps, 3.0 seconds, AAC audio. Three seconds, vertical, with the original audio and movement intact — short, because this is a technique that holds up over seconds rather than minutes.

Vizard Agent checks the motion and the synchronisation on the finished cut rather than assuming the join between segments is clean.

It is not instant, and this category was not separately measured. Comparable work runs a median of 28 to 38 minutes end to end. Across all projects the median cost by tier is Flash 47, Pro 55, Max 242, Ultra 263 credits.

When does this not work well?

The technical limits are narrow and the ethical ones are narrower. Vizard Agent will do this on your own material, and putting a real person's face into footage they did not agree to appear in is not a technical decision.

• Both people have to agree. The one in the photo and the one in the video. This is the limit that matters.
• Never for deception. A real person appearing to say or do something they did not is a fabrication, whatever the intent behind it.
• Label it. Even as a joke, a swapped video circulating without context becomes something else.
• Seconds, not minutes. Quality and cost both work against length. Short clips are what this is for.
• Angles and occlusion break it. A face turning away, a hand across it, motion blur — these are where the swap fails visibly.
• Platforms have their own rules. Several require synthetic media involving a real person to be declared, and Vizard Agent cannot make that declaration for you.

How do you fix a result that came back wrong?

Say what looks wrong and where. Vizard Agent keeps the reference frame, both test versions and each processed segment separately, so trying a different source photograph or reprocessing one segment does not repeat the earlier tests, re-run the other segment, or touch the original audio at all.

• "The likeness is off." A different reference frame or photo, tested on one frame first.
• "The second half looks worse." That segment reprocessed alone.
• "Keep his own clothes." A narrower swap covering the face only.

How does Vizard Agent compare to doing it yourself?

By hand this is a specialist tool, a training step and a long series of trial renders — with no cost discipline at all, because every attempt is a full pass over the whole clip rather than a single frame you can look at and reject in seconds.

Common questions

Do I need permission? Yes, from both people. This is the one requirement Vizard Agent cannot check for you.

How long can the clip be? Seconds. Vizard Agent splits longer clips into segments, and quality and cost both argue for keeping it short.

Does the audio change? No. The original sound is kept and rejoined to the processed picture.

What makes a good source photo? Front-facing, sharp, well lit, nothing across the face. It is the ceiling on the whole result.

Can I test it cheaply first? Yes, and you should. Ask Vizard Agent for a single frame or a three-second test before committing to the full clip.

Will it look real? On a short clip with a clear photo and a face-on angle, close. Turns, blur and occlusion are where it shows.

Does Vizard Agent keep the original movement? Yes. The motion, the scene and the timing are the video's; only the person changes.

Can it swap more than one person? One at a time is far more reliable. Vizard Agent tests each on a frame before processing anything.