Shot Sizes in AI Short Drama: Rules From a 244-Frame Storyboard With 4 Wide Shots
Credits and specs follow TWno1's live billing settings. Last checked: 2026-10-05
Of the 244 storyboard frames in a 16-episode AI drama we reviewed in September 2026, 235 had a person in them and only 4 were wide or empty shots.
No single frame was broken. That is what made it hard to see. A woman in an office, two people talking in a hallway: each frame was fine on its own. Played in order, it was medium shot after medium shot, face after face. Viewers never learned where anyone was standing, and no line landed harder than the one before it.
This piece takes that storyboard apart, gives six rules you can count, and then goes one step upstream to the thing that actually fixed it: a camera setup plan.
The audit
| Measure | Count | Share |
|---|---|---|
| Frames with a person in them | 235 / 244 | 96.3% |
| Wide or empty shots | 4 / 244 | 1.6% |
| Frames with no shot size written on the storyboard | not counted separately | 68% |
| Frames where the script named a shot size and the image ignored it | 69 / 244 | 28.3% |
The rule we later adopted says at least 25% of an episode's frames should have no person, or a very small one. For 244 frames that is 61. This drama had at most 244 − 235 = 9. It missed by 52 or more.
The 68% figure is the root. When the storyboard does not say how wide the frame is, the image model decides. And the most literal picture of "she walks into the meeting room" is a woman in the middle of the frame, cropped at the waist. That explanation is our inference, not something a vendor document states. The 244 frames are consistent with it.
Six rules you can count
"Vary your shot sizes" is advice nobody can check. "The first frame of every scene is a wide shot" is. Every rule below has a number.
| Rule | Threshold | How to check |
|---|---|---|
| Open each scene with geography | At least 1 extreme wide or full shot per scene | Look at the first frame of each scene |
| Leave room to breathe | 25% or more of frames with no person or a tiny one (empty set, prop, hands, detail) | Count frames with and without people after rendering; on the demo drama we ran face and person detection and re-rolled with a new seed when a person showed up, up to 2 times |
| Neighbors differ | Zero pairs of adjacent frames with the same shot size | Walk the shot-size column in order |
| Dialogue gets coverage | Over-the-shoulder plus reaction shots | Check that the listener's face appears in dialogue runs |
| Save close-ups | Close-ups only at turns, reveals, confessions | Match each close-up to a story beat |
| Label everything | Zero frames without a shot size | The column is never blank |
Run the numbers on one episode. The script format we used on the demo drama defaults to 14 frames and at most 3 scenes. So: at least 3 wide or full shots, and at least 14 × 0.25 = 3.5, rounded up to 4, frames without a visible person. Wide shots where the characters are tiny count toward both, so in practice you add one or two prop or hand inserts and you are there.
One trap. Writing "an empty hallway" does not guarantee an empty hallway. Models put people in anyway. Check empty frames with detection after rendering, not by rereading your prompt.
"One line, one shot" is the wrong default
A lot of short-drama advice says to keep things fast by cutting on every line of dialogue. We worked that way. In practice it meant one line equals one frame equals one camera position equals 5 seconds.
Two things went wrong. The picture jumped constantly. Worse, a single location came out looking like three different rooms, because every frame was a fresh camera position and a fresh render, so the furniture and walls got re-guessed each time.
The fix is to plan camera setups before you assign shot sizes:
- Give each scene 2 to 5 fixed setups. For a two-person argument in a kitchen: A, wide from the doorway; B, over his shoulder onto her; C, over her shoulder onto him; D, a tight shot of her for the peak.
- Assign each line to a setup, and reuse setups.
- Every cut needs a reason you can write down. No reason, no cut.
- Narration, inner voice and the other end of a phone call do not get a new picture. They sit over the current shot.
- Duration follows content: short beats around 2.5 seconds, long ones up to 10.
On one 17-line passage, the setup plan produced 13 shots (8 setups plus 2 inserts) running 70.9 seconds. The one-line-one-shot version ran at least 87.8 seconds, so the new cut is at least 16.9 seconds shorter. Unmotivated jump cuts within a setup: zero. Across all 16 episodes, 244 lines became 175 shots.
The setup plan also solves most of the shot-size problem for free. Setup A is a wide. B and C are over-the-shoulders. D waits for the beat. Adjacent frames end up with different sizes almost without trying.
The script said close-up; the image said medium
Of the 69 frames that ignored the script, the typical case was a written close-up that came back as a waist-up medium shot. The rule we adopted is about where the marker sits.
In the demo drama's script format, action lines start with △. The close-up tag has to go at the very start of that line ("△【close-up】her finger stops on the last page of the contract"), not on a line of its own and not at the end. And the storyboard layer must not override a shot size the writer chose. A writer who asks for a close-up wants you to see one small thing.
Vendor docs give you vocabulary that their models are built to read. BytePlus's Seedance 2.0 prompt guide lists medium shot, close-up, wide shot and fixed shot as understood terms, and recommends one camera move per shot. Kling's guides use establishing wide shot, extreme close-up and full body. Google's Veo docs use wide shot, close-up and two-shot. Use their words instead of inventing adjectives.
Limits, and when not to use these rules
- The thresholds are ours, set from one demo. 25% and one wide per scene came out of reviewing these 244 frames. They are not an industry standard. Adjust them by genre, but keep a number.
- There is no solid public number for micro-drama shot length. One single-source industry breakdown we have puts top short dramas at about 2.5 seconds per shot and about 23 cuts a minute; we could not verify it. The firmer research is Cornell's study of 160 Hollywood films, which found average shot length fell below 4 seconds after 2000. Those are horizontal theatrical films. Applying the figure to vertical drama is an assumption.
- Deliberate long takes are exempt. A monologue or a one-shot sequence should not be chopped up to satisfy "neighbors differ." Mark those frames as exceptions.
- Someone has to think. A tool can count shot sizes. Only a writer or editor can say why a cut exists.
Count on stills, not on finished clips. TWno1 draws a storyboard image for every shot and generates video only after you approve it, so "4 of 14 frames without a person" and "every scene opens wide" can be checked while everything is still a picture. If the count comes up short, edit the shot list and redraw those frames; no clip exists yet. To try the count on one episode, open a project in Studio.
Related reading: Vertical Drama Shot Length: Why We Aim for 2.5-4 Seconds and Cap Shots at 6; How Long Does a Mandarin Line Take? A Characters-to-Seconds Table for Short Drama; Scene Plates for AI Drama: Empty Master Shots, 180-Degree Reverse Views and Lighting Variants.
FAQ
How many shots without people should an AI short drama episode have?
Our working threshold is at least 25% of frames with no person or a very small one, such as empty sets, props, hands or detail inserts. For a 14-frame episode that means at least 4 frames (14 × 0.25 = 3.5, rounded up). We set the 25% after reviewing a 16-episode, 244-frame AI drama; it is not an industry standard.
What shot size should open each scene in an AI short drama?
Open each scene with an extreme wide or full shot so viewers know where they are. In the 244-frame AI drama we reviewed, only 4 frames were wide or empty shots, and location changes were confusing, so our rule is at least one wide or full shot per scene.
What is a camera setup plan for a storyboard?
A camera setup plan fixes 2 to 5 camera positions for each scene before any shots are assigned, then attaches each line of dialogue to one of those positions and reuses them. On a 17-line passage it gave us 13 shots running 70.9 seconds, at least 16.9 seconds shorter than cutting on every line, with zero unmotivated jump cuts.
Why does an AI image model draw a medium shot when the script asks for a close-up?
In our review, 69 of 244 storyboard frames (28.3%) did not match the shot size the script asked for. The rule we adopted is placing the close-up tag at the very start of the action line rather than on its own line or at the end, and never letting the storyboard step override a shot size the writer specified.
How do you keep AI drama dialogue scenes from becoming two people in a medium shot?
Cover dialogue with over-the-shoulder shots plus reaction shots, and save close-ups for the emotional peak. If each scene has 2 to 5 fixed camera setups, the rule that adjacent frames must differ in shot size is satisfied almost automatically.
References
- Cutting et al., shot length in 160 Hollywood films (i-Perception, 2011) (checked 2026-10-05)
- BytePlus ModelArk: Seedance 2.0 series prompt guide (checked 2026-10-05)
- Kling: text-to-video prompt guide (checked 2026-10-05)
- Google Gemini API: Veo video generation (checked 2026-10-05)