How to Fix Eye Contact in Any Video
You watch the take back and your eyes are clearly reading — flicking to the side, down to notes, or off toward a second screen. The delivery is fine, but the gaze breaks the connection. Here are the three ways to fix it, from most effort to least.
Only one of them works on footage you have already shot, so if the recording is done, skip to option three.
Option 1: Reshoot (most control, most effort)
The obvious fix is to record again and consciously hold the lens. It works, but it's slow, it rarely survives more than a few takes before your delivery gets stiff, and it's impractical for long recordings or anything you've already published.
It is also the option most likely to cost you something other than time. Holding the lens while recalling what you meant to say splits your attention, and the delivery usually tightens up before the gaze does. If the original take had the energy you wanted, re-recording to fix the eyes often trades one problem for another.
Reshooting is the right call in one situation: when the framing, lighting or audio is also wrong. If the only fault is where you were looking, it's the most expensive way to fix the smallest problem.
Option 2: Use a teleprompter (better, but not free)
A teleprompter that sits directly over the lens keeps your eyes near the camera while you read. It's the pro solution, but a good one costs money, takes setup, and still shows a slight off-axis drift if the script sits beside the lens instead of over it. It also can't help footage you've already shot.
The distinction that matters is beam-splitter versus app. A beam-splitter prompter puts the text on glass in front of the lens, so your eye line and the lens are genuinely the same point. A phone or tablet prompter app parked next to the camera does not do that — it just moves the script closer, which reduces the offset without removing it. That residual few degrees is exactly what viewers read as "looking slightly past me".
Prompters also change how you sound. Reading has a cadence, and the closer you stick to a written script the more the delivery flattens. Many people end up with better eye contact and worse performance, which is not obviously a win.
Option 3: AI eye contact correction (fastest)
AI gaze correction fixes the footage after the fact. You upload the clip, the tool redirects your eyes to the lens frame by frame, and you download a corrected version — no reshoot, no gear. It's the only option that works on video you've already recorded, and it's fast enough to use on every clip rather than just the important ones.
It is now a standard feature rather than a novelty. Descript, VEED, Captions, Kapwing, CapCut and BIGVU all ship a version of it, and NVIDIA's Maxine gaze model sits underneath several of the web tools that offer it. The differences between them are less about the correction itself and more about what you have to buy to reach it — most bundle the feature inside a full editor on a monthly plan.
The trade-off is honesty about limits. Gaze correction re-renders a small region of your face, and on footage it was not designed for, that shows. The next section covers when it fails, because knowing that up front is the difference between a clip you can ship and one you have to reshoot anyway.
Step by step: correcting gaze without a reshoot
Check the preview at full size rather than in a thumbnail. Gaze artifacts are small by definition — a slightly wrong catchlight, an eyelid that moves out of step with the blink — and they are invisible at postage-stamp scale but obvious once the clip is full-screen on someone else's monitor.
- Upload the video (most tools take MP4, MOV, or WebM).
- Preview the before/after side by side to check it looks natural, not robotic.
- Let it process — good tools keep your blinks, look-aways, and audio intact.
- Download the corrected, watermark-free result, ready to cut straight into your edit.
When correction fails, and why
Most tools share the same failure modes, because most of them share similar underlying models. Descript documents its own limits plainly, and they are a fair guide to the category: the effect works on one clearly visible face, will not apply to sideways footage, and can skip sections entirely where the subject turns their head or looks far from the camera.
- More than one face in frame. Descript's documentation states the effect works when one clear face is in frame; several other tools behave the same way.
- Variable frame rate footage. Descript recommends re-encoding VFR video to a constant frame rate before applying the effect. Screen recorders and phones are common VFR sources.
- Large gaze offsets. If your script sat far from the lens, there may be more correction needed than the model will apply, and it can skip those stretches rather than overreach.
- Glasses glare and side lighting. Reflections that hide the eyes give the model nothing to work with. Softer, angled light fixes more of this than any setting will.
- Small or distant faces. Wide shots leave too few pixels in the eye region for a convincing re-render.
Tips for natural-looking results
- Record with your script as close to the lens as possible — less correction means fewer artifacts.
- Shoot at a constant frame rate where you can, and re-encode to CFR before correcting if your recorder produced VFR.
- Keep the face reasonably large in frame; a medium close-up corrects far better than a wide shot.
- Light from the front and slightly above so the eyes stay visible, and angle glasses away from the key light.
- For long videos, correct in one pass rather than clip by clip so the gaze stays consistent.
- Watch the result once at full resolution before you publish. If it reads as a fixed stare rather than someone looking at you, the source footage was probably too far off-axis to rescue.