Case study
Why your AI video presenter grows an extra hand
A client's video team found the trouble with their AI presenter ads was in the hands. Here is where it starts, what we changed, and what to look at in your own clips.
Key takeaways
- A plain ad is forgivable; a strange hand is not, because it breaks the illusion the whole presenter rests on.
- The starting picture decides what the video model has to redraw, so empty, relaxed hands remove the easiest thing to get wrong.
- Ask your video provider what its quality check compares, and whether the hands are part of it.
- Keep earlier takes, so a line can be replaced without losing the version you liked.
- A clip nobody could check should be treated as unverified, never as good.
An AI presenter is a made-up person built from a still picture and a script rather than filmed with an actor, and our client's video team, a business in the United Arab Emirates, uses them for short ads on the client's own platform. In the clips the team had made, the trouble was in the hands: an object that changed as the presenter moved, and a hand where there should not have been one. We changed how those clips are built and how they are checked before an ad reaches the team; the check is live for them, and a faster version replaced it the next day.
What a strange hand costs a viewer
A viewer will forgive a plain ad, but not a strange one, and a single wrong hand is enough for them to stop believing the presenter and move on. The usual offenders are objects: an item in the presenter's hands that turns into something else from one frame to the next, or a phone held by a hand that was not there a moment ago. Until something like that is caught, one bad moment meant producing the whole ad again, which is exactly the time an AI presenter is meant to save.
Why AI presenters get hands wrong
An AI presenter is animated from a starting picture, and the video model redraws whatever that picture puts in the presenter's hands, frame after frame. If the starting photo has him holding something, the model has to keep redrawing that object as he moves, and it does not always redraw it as the same thing. A scene hint written by whoever sets up the shot can ask for a phone, papers or a cup and invite the same trouble. Whether anyone looks at the hands afterwards is a separate question from whether the face is right.
What our engineers found
Going back through the clips, we found the trouble began before any video was made: the presenter's starting photo itself put an object in his hands. The tool's automatic check had passed the clip because it compared the face and the words spoken, and it never looked at the hands. That left two jobs: stop giving the video model things to get wrong, and look at every clip's hands before a person has to.
What we changed
Two things changed, prevention and detection.
- Starting photos, and the still frames each line is animated from, are now made with empty, relaxed hands. Objects are taken out of the scene hints, and every instruction to the video model asks for two hands with natural fingers, nothing picked up and nothing appearing.
- For every presenter clip, four frames are compared with the picture the clip started from, and two yes-or-no questions are asked: is there an extra or malformed hand, and is something in the hands that should not be there?
- A yes sends the clip back to be made once more, like any other failed check. A clip the check could not look at is marked unchecked rather than passed, and every answer is kept, so the team can see why a clip went back.
- The check first looked at one clip at a time, so a run of clips waited their turn. From the next day it looks at three at once while the next clips are still arriving.
- One bad line no longer means a new ad. The team can make a single line again from the editor, say what to fix, and keep the earlier takes to choose from.
When a line's hands failed the check
When a presenter line's hands fail the check, that clip goes back on its own and the reason is kept in the record, so the team can read why it was sent back. The team then remakes that one line from the editor, with a note asking for the hands to be kept still and low and nothing held, and keeps the earlier takes to choose from. Nothing else in the ad is touched, so a second that went wrong no longer pulls the rest of the work with it.
What you can check yourself
- Pause your ad wherever the presenter's hands move, and count the hands and the fingers. An extra hand, a missing finger or a thumb bending the wrong way is what to look for.
- Look at the still picture your presenter was built from. Anything in those hands will be redrawn as the clip moves, and it may change shape or become a different object.
- Watch the closing seconds, where the presenter settles into a pose. A stray hand or a held object can show up there.
- Ask your video provider what its automatic quality check compares, and whether the hands are part of it.
- Ask whether one bad line can be made again on its own, or whether a single wrong moment means producing the whole ad over.
What to do next
For an ad you are about to make, build the presenter with empty, relaxed hands and write the script so he does not pick anything up. For one you already have, watch it once with a finger on the pause button at the places where the hands move, and if something is wrong, ask for that line alone to be made again with the hands kept still and low. Checking what an AI presenter does with its hands, and putting that check in front of every clip, is the kind of work our engineers do for clients.
Questions people ask
Why does my AI presenter suddenly have three hands or extra fingers?
The clip is animated from a starting picture, and the video model redraws whatever that picture put in the presenter's hands, frame by frame. If the starting photo has him holding money, a phone or papers, the model has to keep redrawing that object as he moves, and it does not always keep it looking like the same thing. A check that compares faces and speech can pass a clip whose hands are wrong. Building the presenter with empty, relaxed hands takes the object out of the model's hands before a clip is made.
Can I fix a bad hand without remaking the whole ad?
Yes, if the tool your team uses can remake a single line on its own. Ask for that one line to be made again with the hands kept still and low, and say what needs fixing. Keeping the earlier takes means you still have something to choose from if the new one is worse. If the only option is to produce the whole ad again, it is worth asking the provider whether single-line remakes can be added.
Does an AI video tool check the hands, or only the face and the lips?
That is worth asking your provider directly before you rely on it. The automatic check on the tool behind the clips our engineers went back through compared the face across frames and confirmed the right words were spoken, and it did not look at the hands at all. If your provider cannot say what its check covers, treat the hands as something a person has to review.
Should my AI presenter hold a phone or money in the video?
Anything held in the starting picture has to be redrawn in every frame, and it can change shape or become a different object. Money and phones are the things that went wrong here, and a cup or a stack of papers carries the same risk. If the ad does not need an object in the presenter's hands, leaving them empty and relaxed removes the problem.
What should I ask a video provider before we make an AI presenter ad?
Ask three things: what the automatic quality check compares, whether a clip that fails can be made again automatically, and whether one bad line can be remade without producing the whole ad again. Also ask to see the starting picture before the clip is animated, so you can see for yourself whether the hands are empty.