What it really takes to make a 91-second AI video
We generated 15 cuts and finished a 91-second short with no camera footage. Why one prompt is never enough, and five things to decide before you commission one.
“With AI it only takes one prompt, right?” That is the question we hear most. Here is the actual work behind one 91-second short we made.
Not a single shot came from a camera. The main section is 15 generated clips, followed by a black screen and a closing section, for 91.1 seconds in total. The numbers below were read directly from the editing project file.
Four stages
| Stage | Tool | What happens |
|---|---|---|
| 1. Character sheet | Claude | Lock one person to the same face across all 15 cuts |
| 2. Scene design | Claude | Fix all 15 shots in writing before generating anything |
| 3. Clip generation | Higgsfield | Generate video from the design |
| 4. Finishing | CapCut | Color, subtitles, music and cut lengths |
The first two stages end before a single frame exists. The time spent there is time not spent discarding cuts later.
The face breaks first
The thing that drifts most in AI video is the face. If a different person shows up in each cut, the cuts will not join into one film.
So before any video we produced a sheet that fixes one person from three sides: front, back and a facial close-up. The back view is what keeps the same person in shots taken from behind.
You have to write what not to do
If you only describe what you want, the model drifts a little each time. The sentences that worked best were prohibitions.
- “If the frame turns yellow, the frame is wrong.” When you only ask for a color, each cut gets slightly warmer. Calling it wrong is what makes 15 cuts match.
- “No slow motion” and “only two hands”. Video models slow fight scenes down on their own and add an extra hand.
- “Do not render a horizontal image inside a vertical canvas.” Without it you get black bars above and below even when you asked for vertical.
Generation is material, editing is the film
The generated clips alone did not make a film. All finishing was done by hand.
- 21 video segments: 15 main cuts, the black screen, and 5 cuts re-used from the main section
- 21 subtitle lines: 11 Korean and 10 English
- 4 sound effects and music split into two pieces
- Cut length: generated clips come out at a uniform 5.07 seconds, and only the first three were trimmed to 4.70, 4.80 and 3.50 seconds
We learned one thing. Grade the color once, at the end. We tried to bring in fresh source clips for the closing section and match their color, and it failed. We switched to cutting from the main section, which was already graded.
More cuts were thrown away than kept
We generated far more than the 15 cuts we used. Fingers were off, the camera moved against the plan, or the light direction did not match the neighbouring cuts.
Choosing usable cuts takes most of the working time. The whole process, prompts included, is published in the Hong Kong noir work log.
Five things to decide before you commission one
- Length and ratio. A 15-second vertical and a 60-second horizontal are designed differently.
- Whether a person appears. Making someone resemble a real person needs that person’s consent and face photos.
- One or two reference videos. A mood described in words is heard differently by everyone.
- Subtitle languages. Adding one later means rebuilding the screen layout.
- Where it will run. Each channel has its own safe area and length limit.
Frequently asked questions
Is it cheaper than a live shoot?
It depends. AI has the edge for scenes that need location booking or overseas shooting, and for scenes that do not exist in reality. When the texture of a real product or the face of a real doctor matters, a shoot is the better choice.
Can a clinic use an AI-generated person in an ad?
Clinic advertising has to be checked against medical advertising rules and each channel’s review process first. Whether the ad must be labelled as AI-generated also varies by channel. For specific cases, go through the review body or legal counsel.
How many revisions are possible?
During written design, as many as you need. Once generation is done, changing the person or the background means remaking that cut from scratch. That is why we ask for sign-off at the design stage.
