How to find and fix the shots that give away a generated video
Tell Vizard Agent to clear the errors that make the video look generated. It watches the finished render specifically hunting for morphing, physics that do not hold and visual artifacts, names the shots that fail, and regenerates only those — leaving the edit, the narration and the timing exactly as they were.
What is the short version?
Most people can tell a generated video within two seconds and cannot say why. The reason is usually one or two shots, not the whole film — a hand that reorganises itself, water that falls wrong, an object that quietly changes shape mid-move.
- Go to Vizard Agent with the finished cut.
- Say to scan it for the errors that make it look AI-made.
- Let it regenerate the shots that fail rather than the film.
What do you need before you start?
The finished video, and honestly nothing else. Vizard Agent still holds the prompt and the settings behind every generated shot in the cut, so it can name what went wrong in each one and rebuild only those, without you having to identify any of them by timecode first.
- The cut. As rendered.
- Nothing else, really. It has the shot list.
- Your tolerance. Some tells matter more than others.
- The platform. A phone hides more than a television.
- Time for a second pass. Regeneration is not instant.
What do you type into Vizard Agent?
Ask for the class of problem rather than a specific fix. You are asking Vizard Agent to audit the cut rather than to adjust it, and the useful version of that instruction says out loud what an audit is supposed to be looking for.
Prompt
Variants worth knowing:
- "Tell me which shots fail." The list is useful on its own.
- "Keep the edit." Only the pictures change.
- "Hands and water especially." The two usual suspects.
What does Vizard Agent actually do?
Here is the order Vizard Agent worked in on a real generated short that came back with the tells still in it. Steps one to nine are the normal build; step ten is the audit, and eleven and twelve are the repair.
- Generates the narration and finds a music bed.
- Transcribes its own narration for sentence timings.
- Generates the shots concurrently in the target frame.
- Probes every generated shot to confirm resolution and frame rate.
- Generates the sound effects — ambience, hits, ends.
- Composes the outro card and checks the typography on a frame.
- Generates styled subtitles against the narration timings.
- Assembles the timeline and probes the compiled cut.
- Extracts frames across the timeline and reviews them as a grid.
- Scans the compiled video for artifacts, morphing and physics errors.
- Regenerates the first batch of failed shots.
- Regenerates the remaining failed shots.
Step ten is the one this page is about. It is a different kind of looking from step nine: a frame grid tells you whether the edit works, and it will not show you a hand that changes between frames or a boat that moves in a way water does not allow. That needs the video watched as motion, with the tells named in advance as what to look for.
Steps eleven and twelve are why this is worth doing rather than accepting the cut. Only the failing shots are rebuilt, so the narration, the subtitle timings, the music and the outro are untouched — the fix costs a few shots, not the film.
What does the result look like?
The same film with the giveaway shots replaced. The pacing is identical because Vizard Agent generates the replacements to the same durations, and the subtitles still land exactly where they did, because nothing about the timeline moved underneath them.
What you also get is the list — which shots failed and why. That list is worth reading before you generate the next video, because the same prompt tends to fail the same way twice.
When does this not work well?
An audit finds the shots that break, but it cannot make a generator do something the generator cannot do. Some subjects fail reliably no matter how many attempts they are given, and knowing which ones those are in advance saves you a long afternoon of regenerating the same shot.
- Close-up hands doing fine work. Still the hardest thing.
- Readable text inside the scene. Signs, labels, screens.
- Complex crowd motion. People pass through each other.
- A specific real person. Likeness drifts between shots.
- Long continuous takes. The longer the shot, the more drift.
How do you fix a result that came back wrong?
Name the shot and what you saw. Vizard Agent keeps every shot's prompt, the timeline and the audit findings, so a second attempt at one shot slots back into the same slot at the same length without disturbing anything around it.
- "The hand is still wrong." Regenerated, or reframed to avoid it.
- "Now it does not match." Re-graded against its neighbours.
- "Cut that shot entirely." Removed; the gap re-covered.
- "Use a still instead." A held frame, which never morphs.
How does Vizard Agent compare to doing it yourself?
By hand this means watching your own video ten times until you stop seeing it, then asking someone else, then regenerating shots one at a time and re-editing around whatever length comes back. The watching is the hard part — you lose objectivity on your own cut within about three viewings.
| By hand | Vizard Agent | |
|---|---|---|
| Spotting the tells | Your eye, after too many viewings | Scanned for named failure types |
| Knowing which shot | Scrub and guess | Listed by the audit |
| Regenerating | One at a time, manually | Batched, at the same durations |
| Re-editing after | Timings shift | The timeline never moves |
| The next video | Same mistakes | The failure list tells you why |
Common questions
Can it check a video it did not make? Yes. Vizard Agent watches any upload for the same failure types.
Will it catch everything? No. Vizard Agent reports what it finds; a second pair of eyes still helps.
Does regenerating change the length? No. Vizard Agent rebuilds to the same duration.
Can I ask for just the list? Yes. Ask it to report and not to change anything.
What are the most common tells? Hands, text in scene, water, and objects changing shape mid-shot. Vizard Agent looks for all four.
Can it fix a bad shot without regenerating? Sometimes — a reframe or a shorter hold hides it. Vizard Agent will say.
Does this cost as much as the first pass? No. Only the failed shots are regenerated.
Can it scan for one specific problem? Yes. Name it and Vizard Agent looks for that.
Will the replacement match the grade? Yes. Vizard Agent grades it against the surrounding shots.
How many attempts does a shot get? As many as you want, but some subjects will not come good.
Does it check the audio too? If you ask. Vizard Agent measures loudness and sync in the same pass.
Can I approve each replacement? Yes. Ask to see them before they go into the cut.
Should I just film it instead? For hands, text and specific people — often yes, and Vizard Agent will tell you so.
Why does the first pass miss these? Because generation and inspection are different jobs. The audit is a second look with a specific brief, which is exactly why it finds things the build did not.