Vizard Agent

How to fix eye contact in a talking-head video

Last updated 2026-08-16 · 6 min read

Upload the clip to Vizard Agent and say the speaker is looking away from the camera. Vizard Agent checks the correction is actually available, extracts frames to see where the eyes are pointing, processes the segment that needs fixing, and leaves the rest of the footage exactly as you filmed it.

What is the short version?

Someone reading from a script below the lens looks like they are avoiding you. The delivery is fine, the audio is fine, and the only problem is a few degrees of gaze — which is not worth a reshoot and used to be the only fix.

  1. Go to Vizard Agent and upload the clip.
  2. Say the speaker is looking down or away and should be looking at the camera.
  3. Say plainly that nothing else about the video should change.

What do you need before you start?

The clip and one instruction. Vizard Agent inspects the footage to see how far off the gaze actually is, so what it needs from you is the boundary — that this is a correction rather than an invitation to restyle the whole video.

What do you type into Vizard Agent?

State the problem and the constraint in one sentence. Vizard Agent extracts frames and judges the gaze itself rather than needing it measured, so the genuinely useful thing you add is the instruction to leave everything else about the video exactly as it was filmed.

Prompt

Fix the eyes so [the speaker] is not looking down the whole time, without changing anything else about the video.

Variants worth knowing:

What does Vizard Agent actually do?

Here is the order Vizard Agent worked in on a real gaze correction. The pattern worth noticing is at both ends: it confirms the capability exists before planning around it, and it keeps trying routes when a step fails rather than reporting a dead end.

  1. Checks which tools are available for correcting gaze direction.
  2. Probes the clip for duration and technical details.
  3. Reads how the video-editing tool works for this specific correction.
  4. Extracts frames to see the face and where the eyes are pointing.
  5. Writes its own frame-extraction script, then fixes it when the first version is wrong.
  6. Looks at the extracted frames to confirm the gaze problem is real and how severe it is.
  7. Trims the segment that needs processing rather than sending the whole file.
  8. Uploads it and starts the correction.
  9. Diagnoses the failure when the upload path is rejected, reads the log, and works through alternative routes until it goes through.

Step 9 is the honest part of this page. The run hit a real obstacle in the middle, and Vizard Agent read the error, understood what had gone wrong and found a way around it rather than handing the problem back.

What does the result look like?

From the run this page is written from, probed on the delivered file: 1080x1920, H.264, 25fps, 25.05 seconds, AAC audio. Vertical, twenty-five seconds, the same framing and the same audio as the original, with the gaze corrected and nothing else altered.

Vizard Agent processes the section that needs it rather than re-encoding the whole file, so what you get back is your video with a change in it, not a new version of it.

That distinction is the whole point of this job. Vizard Agent treats the original as the thing being preserved and the correction as the smallest possible intervention in it.

How long it takes was not measured separately here. Comparable work runs a median of 28 to 38 minutes end to end, and the upper quartile is several times the median. Across all projects the median cost by tier is Flash 47, Pro 55, Max 242, Ultra 263 credits.

When does this not work well?

Gaze correction is a subtle change to a face, and faces are what people look at hardest. Vizard Agent applies it carefully to the section that needs it, and there is a point past which the correction is more distracting than the original problem.

How do you fix a result that came back wrong?

Say what looks off. Vizard Agent keeps the original, the extracted frames and the corrected segment separately, so backing the effect off or applying it to a shorter stretch does not touch the audio or re-encode the parts that were already fine.

How does Vizard Agent compare to doing it yourself?

By hand this is either a reshoot or a plugin in an editor that you have to find, licence and learn — for a problem that shows up in one clip out of ten and is never worth setting up for in advance.

By hand Vizard Agent
Knowing it is possible Search for a plugin Capability checked first
Judging the gaze Watch it back Frames extracted and examined
Processing The whole file Only the section that needs it
A failure mid-way Give up or start again Log read, another route found

Common questions

Will it change the rest of my video? No. Vizard Agent processes the affected section and leaves the remaining footage as you filmed it.

How far off can the gaze be? Small angles work well. Someone reading from a desk well below the lens is beyond what a correction can fix convincingly.

Does the audio change? No. This is a picture correction and your original sound is untouched.

Can I see it before committing to the whole clip? Yes. Ask Vizard Agent to correct a short section first and look at that before it processes the rest.

Can it fix only part of the clip? Yes, and that is usually better. Name the seconds where the gaze is wrong.

Can it do skin and colour at the same time? Yes. Ask for them together and Vizard Agent inspects the footage once instead of twice.

Is it obvious? Not at feed size when the angle was small. Check it at full size before publishing anything that will be seen large.

Does it work on more than one person? Vizard Agent corrects the speaker in frame. Two people at different angles is a harder case and worth checking on a short section first.

Will the video be re-encoded? Only the part that changes. Vizard Agent leaves the untouched footage as it was rather than rebuilding the whole file.