Storyboard
When to draw one frame per shot and when to make one storyboard sheet, how to feed a sheet to Seedance or Omni, and prompts you can copy.
A storyboard turns your shot list into pictures before you pay for video. Images cost far less than video, so mistakes are cheap to fix here. For a single clip under about 15 s, a one line note per shot is enough. For anything longer, or with more than one character, board the key beats.
One sheet or one frame per shot
You have two options. Use the sheet to plan, then make single frames only for the shots that need them.
| One storyboard sheet | One frame per shot | |
|---|---|---|
| What it is | One image with 6 or 9 panels, one panel per beat | A separate full size image for each shot |
| Cost in the JXP test (6 shots, GPT Image 2.5) | 1 credit | 6 credits |
| Time in the same test | 32.5 s | 3 min 13 s |
| Continuity | Props, palette and character hold across panels | Broke: a ceramic cup became a paper cup, the espresso machine changed colour |
| Resolution | Low: at 1K, about 450x380 px per panel for 6 panels, about 450x250 px for 9 | Full size, usable as a first frame |
| Best for | Planning, review, and feeding the whole sheet to the video model | The 2 or 3 hero shots you animate with image to video |
These numbers come from one careful test by JXP. Plan on the sheet, then rebuild only the 2 or 3 keeper frames at full resolution. If one panel is wrong, regenerate the whole sheet (1 credit) rather than patching a single panel.
The storyboard sheet trick
In 2026 many creators skip writing long video prompts. They make a grid storyboard with an image model and give the sheet to the video model as the prompt. Google's Omni guide supports this officially, and creators on X report the same pipeline with Seedance: script, character sheet, 3x3 storyboard, video.
Lock the character first
Make a character sheet before the storyboard and attach it to the storyboard prompt. A wrong detail on the reference (a shoe colour, for example) repeats in every panel. See Reference images.
Generate the sheet
Use GPT Image 2.5 or Nano Banana (Nano Banana Pro or Nano Banana 2). Put the grid layout and panel count at the very start of the prompt, or you get one big image instead of a grid. Kapwing used 15 panels for a 30 s video (about 2 s per panel), with the sheet at 4K and 16:9.
Review it like a director
Check each panel against the shot list: one framing, one action, same character, same light. Fix the prompt and regenerate the whole sheet until it reads as a story.
Feed the sheet to the video model
Attach the sheet to Seedance 2.5 or Gemini Omni with a short instruction: the reading order, the total duration, which image holds the character, and "no text". Omni's official guide asks you to state the reading order and total duration.
Expect beats, not exact frames
The video model treats each panel as a beat. Style and character carry over, but exact layouts do not. If a shot must match a precise composition, rebuild that panel as a full size first frame and use image to video.
Aspect ratio
Set the aspect ratio yourself. On "auto", image models copy the shape of the reference image.
- 16:9 for YouTube and standard video.
- 21:9 for a widescreen, anamorphic look.
- 9:16 for TikTok, Reels and Shorts. Either make a 9:16 sheet, or a 16:9 sheet that holds vertical panels.
Decide this once, before any keyframe, so every frame matches the final video.
What to put in a storyboard prompt
- The grid first: "3 columns x 2 rows, 6 equal panels".
- A continuity lock: face, hair, wardrobe, location, lighting direction, palette, recurring props.
- One shot type and one action per panel.
- The style line again, so panels do not drift toward photoreal.
- Scoped negatives. "No extra characters" on its own emptied a café that was meant to be busy. Write "no extra characters except where a panel describes them".
- What is visible in the frame, not where the camera stands. "From behind the counter" was followed less reliably than describing what we see.
- Panel numbers outside the panel art, plus "no text inside panels". Labels inside panels show up in the video.
Text leaks into the video
If panel numbers, captions or borders appear in your clip, move the labels outside the panels and add "Do not render any text, numbers, panel borders or captions" to the video prompt.
Example prompts
A 6 panel sheet for a 16:9 story (GPT Image 2.5 or Nano Banana):
A storyboard sheet: 3 columns x 2 rows, 6 equal panels with thin white
borders, panel numbers 01 to 06 in the margin outside each panel, 16:9 sheet.
Identity reference: the attached character image is MAYA; keep her face,
hair, glasses, mustard apron, white tee and proportions identical in every panel.
Continuity lock: small corner café, warm morning window light from the left,
palette cream, teal and brass, the same white ceramic cup throughout.
Style: soft painterly illustration, muted palette, not photographic.
01 wide: Maya opens the café door, the café is empty.
02 medium: Maya ties her apron behind the counter.
03 close up: espresso pours into the white ceramic cup.
04 over the shoulder: Maya slides the cup to a customer (the only extra person).
05 close up: the customer smiles, steam rising from the cup.
06 wide: the café is now busy, Maya smiling behind the counter.
No text inside the panels.A 9 panel sheet for a vertical short:
3x3 grid of 9:16 panels on a single sheet, numbered outside the frames,
reading order left to right, top to bottom. One character: <character bible>.
One location: <location bible>. Each panel is one beat, one framing, one action:
panel 1 is the hook, an extreme close up of a strange detail; panel 9 is a
loopable ending that matches panel 1. Same lighting direction and palette
in all panels. No captions.Then the instruction that goes with the sheet into the video model:
Seedance refers to uploads as @Image 1, @Image 2 in upload order. Say what each one is for.
Create a 30 second video from @Image 1 (the storyboard). Follow the panels
strictly in reading order, about 2 seconds per panel, do not skip panels or
finish early. Character identity from @Image 2 (character sheet) only.
Do not render any text, numbers, panel borders, timecodes or captions.
Audio: ambience and foley only, no music.Omni makes up to 10 s per step and extends to 40 s, so for a longer board split it into parts. Seedance 2.5 makes up to 30 s in one generation. See Seedance and Google.
Planning camera moves in 3D (optional)
Some creators block the scene in Blender or another 3D tool first and report that Seedance then follows the camera closely and needs fewer rerolls. Pippit 3D Director Studio offers this as a product (place characters, preview pans, dollies and orbits, keyframe a timeline), but the only sources are Pippit's own pages, so treat it as unreviewed. For most beginners the 2D sheet is enough.
Sources
- Sheet versus single frames, panel sizes, prompt rules: JXP
- 15 panel workflow, labels leaking: Kapwing
- Omni storyboard input: Google DeepMind Omni prompt guide, Gemini API Omni docs
- "The storyboard is the prompt": Pletor
- When to storyboard: invideo FAQ
- Creator pipelines and 3D blocking on X: beginnersblog1, EHuanglu, rewind02
- Seedance reference syntax and durations: BytePlus ModelArk, Seedance 2.5
- Pippit: Pippit 3D Director Studio