Case study
Why your AI video passes every check and still looks wrong
An AI video can pass every automatic check and still put the presenter where no person could sit. On a client's platform, one such picture spoiled the second half of a sales video; here is what we found and changed.
Key takeaways
- Automatic checks can prove shape and face, but they do not prove that the scene makes sense.
- A single model's overall opinion is unreliable; narrow questions with fixed rules are steadier.
- Check the picture every later scene is built on before filming starts.
- Test a new check on past results, including ones you know went wrong.
- Put the strongest check where a mistake would spoil the most work.
When an AI video looks wrong halfway through, the cause is often a single picture that no automatic check questioned. On a client's platform, a sales video moved the presenter into a vehicle halfway through, and the picture drawn for that place put him where no person could sit. The fix was live the same day: new pictures are now checked before any video is filmed from them, and an impossible one is caught and redrawn.
What a viewer sees that your checks do not
A viewer spots a person floating where no seat is, even when every automatic check has passed. In the video, the presenter moves to a second place halfway through: a seat inside a vehicle. The picture drawn for that place put him between the front seats, over the middle of the dashboard, with his head against the roof. Every line after the change was filmed from that one picture, so the second half carried the mistake. A viewer may forgive a small flaw, but not a body in a place it cannot be. The team usually finds out by watching the finished result, after the work is done.
Why an AI video looks wrong after a change of scene
The picture passed every automatic quality check because each check measured identity and framing, not whether the scene was physically possible. The platform checked the shape of the image, looked for black bars, checked that the mouth sat at the right height for lip movement, and compared the face with the presenter's photo. Each check did its job. None asked whether the person was sitting where a person can sit. The instruction for drawing a new place also never said how a person is placed in it, so the AI tool that draws the picture filled the gap as it saw fit.
The detective work: from a model's opinion to plain questions
The initial attempt was a quick, inexpensive vision model (an AI tool that looks at a picture and answers questions about it) asked for an overall opinion of each new place. It was not reliable: shown the same vehicle picture twice, it passed it once and failed it the next time. An opinion that changes from one look to the next cannot be the only guard on a long video. The turning point was to stop asking for an overall opinion and ask a few plain yes-or-no questions instead. Fixed rules of our own then turned those answers into a pass or a fail.
What we changed
Every picture drawn for a new place is now reviewed before any video is filmed from it, once the face is confirmed as the presenter's. The vision model answers six yes-or-no questions. Inside a vehicle: is he in a seat with its backrest behind him, and never between the seats? Anywhere: is there an impossible pose, size or body? Our rules turn those answers into a pass or a fail. Inside a vehicle the stronger model decides; elsewhere the quick reader's pass stands, and any fail it gives is checked again by the stronger one. A failed picture is drawn once more with the pose at the front of the instruction, then dropped if it fails again, so it never reaches the video. The drawing instruction now says how the presenter is placed in every new place: in a vehicle, in the seat the script names, otherwise the driver's, with the belt on and the view from the dashboard.
What the business has now
The check is live on the client's platform, so a picture in an impossible position is caught and redrawn before filming starts. Before it went live, the new check ran on about twenty second-place pictures from earlier videos: every vehicle picture failed, and every other one passed. New vehicle pictures drawn with the new instruction passed too.
What you can check yourself
- Watch the opening seconds after every change of scene, not only the start of the video. A presenter between the seats, a body at the wrong size or a pose no person can hold is the kind of fault to look for.
- List what your automatic checks actually test. Then ask what a viewer would spot that none of them covers.
- Show the same image to your AI checker twice. If the verdict changes, it cannot be trusted on its own.
- Ask whether one bad picture can spoil everything built after it. If it can, check that picture before the rest.
- Play the finished video at normal speed on a phone. A fault that is hard to see in a still is often plain in motion.
What to do next
Before your next AI video, review the join after each scene change and ask whether a person could really be there. If the answer is no, stop before the filming step and redraw the picture with the pose made explicit. This is the kind of check our engineers build for clients who make videos with AI, and it is worth testing on an old result you already know was wrong.
Questions people ask
Why did my AI video look fine at the start and wrong halfway through?
The opening scenes were built from one picture, and the later scenes from another. If the picture for a new place is physically impossible, every shot made from it carries the same fault, even when the checks on face, shape and framing pass.
Can I trust an AI model to judge whether a generated picture looks right?
Not as a single judge giving an overall opinion. A model can answer narrow yes-or-no questions, but it may give a different verdict on the same picture, so fixed rules should decide and any fail should be checked again.
How do I stop a bad AI picture from spoiling the rest of a long video?
Review each new scene picture before any filming begins. If it fails, redraw it once with the pose made explicit, and drop it if it fails again rather than letting it through.
Where should I look when reviewing an AI video before publishing?
Start with the opening seconds after every change of scene, because that is where a new picture takes over. Then play the rest at normal speed; a wrong seat, a floating body or an impossible size is easier to see in motion than in a still.