AI Image & Video
Stop AI Drawing the Wrong Person in 3x3 Storyboards: Prompt Template and Checks
Nine panels crammed into one image and one prompt leave the model unable to tell who 'she' is. Replace every pronoun with a fixed character code and the usable rate went from 33% to 75%.
The one idea that matters
Nine panels share one image and one prompt, so the model cannot tell who a pronoun points to. Swap she, he, they, the two and both for fixed character codes and most errors disappear.
Steps
Step 1: Define the characters first (write once, never change a word)
Use ALL-CAPS English codes. Describe features that make gender readable, and give the two characters different clothing colors:
THE WOMAN = a young East Asian woman in her mid 20s, hair in a low bun, wearing a dark indigo robe
THE MAN = a young East Asian man in his late 20s, short black hair, wearing a dark charcoal robe
Step 2: Build the prompt from this skeleton (keep the order)
A storyboard sheet: a strict 3x3 grid of nine equal-sized 16:9 panels separated by thin white gutters, three rows and three columns, nine panels total, no text, no numbers, no captions.
Style for every panel: <style>. Setting for every panel: <setting>.
Recurring characters, identical in every panel: THE WOMAN = <look>. THE MAN = <look>.
Panel 1 (top-left): <shot size>, THE WOMAN stands near the doorway; exactly one person: THE WOMAN, no other people.
Panel 2 (top-center): <shot size>, THE WOMAN (left) and THE MAN (right) sit at a table; exactly two people: THE WOMAN and THE MAN, no other people.
...continue through Panel 9 (bottom-right)
Position names for the nine panels: top-left, top-center, top-right, middle-left, center, middle-right, bottom-left, bottom-center, bottom-right.
Step 3: Run these four checks before sending
- Does any panel contain he, she, they, the two or both? Replace with the character code (“both hands” does not count).
- Do panels with two or more people carry (left) / (right) positions?
- Do two-person panels use his, her or their? Replace them, since it is unclear who they refer to.
- Are there exactly 9 panels?
Step 4: Generate 4, pick 1
Check panel by panel: right head count, right genders, hair and clothes consistent for each character, no stray text. At a 75% pass rate, the chance that all 4 fail is under 1%.
Traps you will hit
- “Two people, front view, wide shot, sitting side by side” fails most often. Small faces plus same-color clothes make the model draw both as the same gender. Restate the looks in that panel:
THE WOMAN (left, hair in a bun) and THE MAN (right, short hair). If it still fails, switch to a side view or over-the-shoulder shot. - Do not feed the previous 3x3 sheet back as a reference to keep characters consistent. The faces lock, but the model copies the reference’s composition (a rainy-night farewell turns back into daytime tea) and the woman-turns-man error comes back. A single-person portrait as a reference also breaks the 3x3 grid. For consistency across sheets, the only method that works so far is “character definitions unchanged to the word” plus picking the best of several.
- Once you turn the sheet into video, the opening flashes the grid. The image-to-video model’s frame 0 is the collage itself, and the leftover length varies (0 to 1.3 seconds in our tests), so cutting a fixed number of frames is unreliable. Detect white divider lines at the 1/3 and 2/3 positions frame by frame and trim up to the last grid frame.
- After crossfading clips with ffmpeg, the file will not open. The crossfade filter promotes video to H.264 4:4:4, which Windows players and phones do not support. Add
-pix_fmt yuv420pon output.
Prompts to paste into an AI
When you are not sure how to write the nine panels in English:
I am a beginner and want an AI to generate a 3x3 cinematic storyboard sheet. Here are my nine panel descriptions: [paste nine lines]. Characters: [characters and looks]. Please rewrite them as an English prompt with these rules: no he, she, they, the two or both in any panel, only ALL-CAPS character codes (for example THE WOMAN); mark (left) / (right) in panels with two or more people; end every panel with “exactly N people: …, no other people”.
When you are stuck:
I am a beginner using [your tool] to generate images. My prompt is: [paste]. The result is: [describe which panel is wrong and how]. Walk me step by step through the likely cause and which sentence to change.
Takeaway
Replace the “she” in your prompt first. Then talk about rerolling.
This is part of an ongoing series of field notes and AI tool tests. To get the full SOPs and future posts, subscribe to the newsletter.
Support this work:Support on Ko-fi