A story told in panels fails for the same reason a scene fails: the beats were never separated. Ask an image model for “a comic page about a chase” and you usually get one crowded picture with three half-actions inside it.
This guide covers the brief: what to decide before you prompt, which documented controls map to which decision, and how to repair one panel without losing the sequence around it.
Split the story into beats before you prompt
A beat is a change, not a moment. “Mara runs” is a state; “Mara finds the door locked” is a beat.
- Write the beats as a numbered list in prose first, with no image words in it
- One beat becomes one panel; a beat that needs two panels was two beats
- Six to nine panels is a workable session; beyond that the later frames drift
- Name what changes in each beat: position, knowledge, or who has the advantage
Do this on paper. It decides whether everything after it works.
Decide whether you are making a storyboard or a comic page
One keyword covers two different products.
- A storyboard panel carries rhythm and camera. It can be rough, and it usually has no lettering
- A comic page adds layout, panel borders, balloons and reading order
- The page is downstream of the storyboard: you cannot place balloons on beats you have not separated
- Both are raster images, not layered editable files, and no OpenAI documentation claims otherwise
Decide now, because the page brief is longer and mixing the two wastes generations.
Write one panel at a time
Six fields per panel, written as a list keyed to the panel number.
- Shot — “wide”, “medium”, “close on hands”
- Subject — who is in frame, named the same way every time
- Action — one verb, one beat
- Place — the same location described with the same nouns throughout
- Light — time of day and direction, held constant until the story changes it
- Exclusions — no lettering, no extra characters, no border
Anything unspecified gets filled in for you, and two actions in one panel is the commonest cause of a muddy frame. Adding a second figure to an existing frame is a separate problem, covered in combining images in GPT Image 2.5.
Fix the frame with ratio, distance and sketch
Three documented controls do the composition work.
- GPT Image 2.5 generates images in any aspect ratio: use the aspect ratio picker, or state the ratio in the prompt
- State it anyway, so a vertical strip and a wide panel cannot be confused for each other
- Sketch lets you draw in ChatGPT as a reference for the final image, the cheapest way to show where figures sit — see drawing with Sketch in GPT Image 2.5
- Sketch is a mobile-app feature; on the web, describe the layout in words
Camera vocabulary beats adjectives. “Low angle, close on hands” fixes a shot that “dramatic” never will.
Hold the same character across panels
One panel is easy; six panels that read as one character is the job.
- Generate the first panel, then feed it back as a reference for the next one instead of describing the character again
- OpenAI states the improvement directly: the model is better at working from reference photos, and distinctive features are more likely to carry through
- Change one line of the prompt per panel, so a mismatch is easy to spot
- Describe wardrobe and build with the same words every time, because wording drift shows up as a different person
- This is the same problem as keeping a subject recognisable, which we walk through in holding one character across images
Two worked examples: a chase and a conversation
Both use the six fields above. Run them as written, then change one line at a time.
- Three-beat chase (2:3 vertical). Panel 1: “Wide shot, 2:3 vertical. A courier in a red jacket sprints through a night market alley, seen from behind, low angle, hanging bulbs and wet ground. One figure, no text.” Panel 2: “Same alley, same red jacket, medium shot from the side as they vault a stall, same bulbs, no other people, no text.” Panel 3: “Same alley and jacket, close on a hand pressing a door handle, same lighting, no text.” Check in this order: the same jacket and build in all three, the same door positions, and only one action per panel.
- Two-panel conversation (16:9). Panel 1: “Wide shot, 16:9. Two people sit across a small cafe table, silhouetted against a bright window, two cups on the table, no text.” Panel 2: “Close shot, 16:9. The same table, tight on two hands and the cups, same window light from behind, no faces, no text.” Check whether the table and cups match panel 1, and whether panel 2 reads as the same moment rather than a new scene.
Lettering belongs in the page layout, not in the image.
Repair one panel without redrawing the sequence
Regenerating a whole strip to fix one frame is how a sequence loses its character.
- Comments can be placed directly on an image for more focused editing, the documented way to point at one region
- The second route is to describe the change in the conversation panel without the selection tool
- OpenAI states the limit plainly: highlights are not always precise and edits may extend beyond the area you selected, so check the frames on either side
- Multi-turn consistency is a documented improvement: earlier changes are more likely to stay consistent as you keep editing
The full workflow is in comment editing in GPT Image 2.5.
Where comic storyboards still break down
Plan for these, because no prompt removes them.
- Repeated props drift: check the same alley or table panel by panel, not as a group
- Lettering inside the frame: the more words you add, the likelier one is misspelled, and no accuracy figure is published
- Complex camera moves: a shot carrying two ideas gets averaged, so split it into two panels
- Crowd panels: many figures in one frame is where faces and hands stop holding up
The honest workflow is two passes: beat the story and generate, then correct the panels. Control details are in the GPT Image 2.5 overview.
Frequently asked questions
Can GPT Image 2.5 generate a whole comic page in one prompt?
It can generate a page-shaped image, but that is one picture rather than a laid-out page with readable panels and balloons. Splitting the story into beats is what makes a sequence hold up.
How many panels should I plan for?
Six to nine per session is workable. Beyond that the later panels drift from the earlier ones, and you spend more time repairing than generating.
How do I keep the same character in every panel?
Generate the first panel, feed it back as a reference for the next, and change only one line of the prompt per panel. OpenAI states the model is better at preserving subjects from reference photos.
Can I fix one panel without regenerating the others?
Yes. Mark the region with the selection tool or comment on the image, describe the correction, then check the neighbouring panels, because OpenAI notes that edits may extend past the selected area.