How do you create consistent stories in images with GPT4o?

GPT4o can now generate pretty good images when given proper instructions.

However, it’s still hard to get consistent images between successive generations.

If you need continuity between different images, either in the same chat, or in different windows, here’s a tutorial to get good consistency across different runs.

1. Storyboard generation

For this quick experiment, I’m going to directly ask ChatGPT to create a story.

Several “frameworks” can be used to generate stories (“The Hero With a Thousand Faces”, from Joseph Campbell, “Save the Cat” from Snyder, “The Shape of Stories” from Kurt Vonnegut, etc…).

Here’s the prompt (I used “o4-mini-high” here):

Create a compelling story, following the framework "The Hero with a thousand faces", or the one from Vonnegut. I want a series of scenes, like a storyboard.

Here’s a partial screenshot of the answer:

Partial screenshot of ChatGPT story generation response

Full result:

Scene	Stage	Visual Description	Key Action / Conflict	Dialogue / Caption
1	Ordinary World	A dusty workshop lit by lanterns. Shelves of half-drawn maps.	Mira, a young cartographer’s apprentice, traces a faded star map.	Caption: “Every day, I chart lands I’ll never see…”
2	Call to Adventure	A wizened traveler bursts in, clutching a glowing compass shard.	He begs Mira: “Only you can restore the Celestial Compass!”	Traveler: “Without it, the world will fall to darkness.”
3	Refusal	Mira recoils, sketchbook trembling in her hands.	Fear grips her: she’s never left the safety of the workshop.	Mira (thought): “I’m no hero—I’m just an apprentice.”
4	Meeting the Mentor	By moonlight beneath a grand world map, the Traveler teaches Mira basic rites.	He reveals ancient lore: compass shards are scattered across three realms.	Traveler: “Your maps will guide you where magic won’t.”
5	Crossing the Threshold	Mira steps through an archway of swirling starlight into a barren desert plain.	The boundary shakes; her mentor’s voice echoes: “Trust your hand.”	Caption: “And so my map became my sword…”
6	Tests, Allies, Enemies	Montage of three trials: defeating sand wraiths, befriending a desert fox, solving a living labyrinth of rock.	Mira learns to read the desert’s hidden signs.	Fox (gesture): Points to a glowing waypoint on her map.
7	Approach to the Inmost Cave	Night-soaked forest. A massive, gnarled tree pulses with dark energy.	Mira confronts her doubt as the forest whispers her failures.	Caption: “Every step forward echoed my worst fears…”
8	Ordeal	Inside the tree’s hollow, a shadow-twin of Mira emerges, mirror-staring her.	She must overcome self-doubt to claim the second compass shard.	Shadow-Mira: “You’re not worthy.”
9	Reward (Seizing the Sword)	The shadow dissolves; Mira holds the glowing shard aloft, light spilling outward.	Confidence blooms—she’s no longer the timid apprentice.	Caption: “Light follows those brave enough to stand.”
10	The Road Back	Snowy mountain pass. Ancient golems awaken, chasing her toward the final realm.	With two shards in hand, she races time to the Temple of Stars.	Mira (panting): “Just… one… more…”
11	Resurrection	At the summit, under a blood-red eclipse, Mira faces the final guardian—a celestial dragon.	She must sacrifice her only complete map to restore the compass.	Dragon (roar): Ember sparks illuminate her map’s sacrifice.
12	Return with the Elixir	Back in the workshop at dawn, Mira rebuilds her maps—now alive with starlight.	The fully restored compass guides ships and scholars worldwide.	Caption: “In giving everything, I found my true path.”

The story has 12 beats. The pace of the story’s tension is outlined in the following graph (I asked Claude to create a SVG from the table directly). The following prompt was the one used with Claude:

Here's a table summaruzing the 12 beats of a story : "| Scene | Stage | Visual Description | Key Action / Conflict | Dialogue / Caption | | ----- | --------------------------- | -------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------- | --------------------------------------------------------------- | | 1 | Ordinary World | A dusty workshop lit by lanterns. Shelves of half-drawn maps. | **Mira**, a young cartographer’s apprentice, traces a faded star map. | *Caption:* “Every day, I chart lands I’ll never see…” | | 2 | Call to Adventure | A wizened traveler bursts in, clutching a glowing compass shard. | He begs Mira: “Only you can restore the Celestial Compass!” | **Traveler:** “Without it, the world will fall to darkness.” | | 3 | Refusal | Mira recoils, sketchbook trembling in her hands. | Fear grips her: she’s never left the safety of the workshop. | **Mira (thought):** “I’m no hero—I’m just an apprentice.” | | 4 | Meeting the Mentor | By moonlight beneath a grand world map, the Traveler teaches Mira basic rites. | He reveals ancient lore: compass shards are scattered across three realms. | **Traveler:** “Your maps will guide you where magic won’t.” | | 5 | Crossing the Threshold | Mira steps through an archway of swirling starlight into a barren desert plain. | The boundary shakes; her mentor’s voice echoes: “Trust your hand.” | *Caption:* “And so my map became my sword…” | | 6 | Tests, Allies, Enemies | Montage of three trials: defeating sand wraiths, befriending a desert fox, solving a living labyrinth of rock. | Mira learns to read the desert’s hidden signs. | **Fox (gesture):** Points to a glowing waypoint on her map. | | 7 | Approach to the Inmost Cave | Night-soaked forest. A massive, gnarled tree pulses with dark energy. | Mira confronts her doubt as the forest whispers her failures. | *Caption:* “Every step forward echoed my worst fears…” | | 8 | Ordeal | Inside the tree’s hollow, a shadow-twin of Mira emerges, mirror-staring her. | She must overcome self-doubt to claim the second compass shard. | **Shadow-Mira:** “You’re not worthy.” | | 9 | Reward (Seizing the Sword) | The shadow dissolves; Mira holds the glowing shard aloft, light spilling outward. | Confidence blooms—she’s no longer the timid apprentice. | *Caption:* “Light follows those brave enough to stand.” | | 10 | The Road Back | Snowy mountain pass. Ancient golems awaken, chasing her toward the final realm. | With two shards in hand, she races time to the Temple of Stars. | **Mira (panting):** “Just… one… more…” | | 11 | Resurrection | At the summit, under a blood-red eclipse, Mira faces the final guardian—a celestial dragon. | She must sacrifice her only complete map to restore the compass. | **Dragon (roar):** Ember sparks illuminate her map’s sacrifice. | | 12 | Return with the Elixir | Back in the workshop at dawn, Mira rebuilds her maps—now alive with starlight. | The fully restored compass guides ships and scholars worldwide. | *Caption:* “In giving everything, I found my true path.” | ". Create an image with an arrow illustrating the pace of the story, its tension, going from left to right, up and down. The prompt used to create this table was : "Create a compelling story, following the framework "The Hero with a thousand faces", or the one from Vonnegut. I want a serie of scenes, like a storyboard.". Create the image. Create a SVG, instead of an image

The resulting SVG, exported as PNG:

Story arc graph showing tension across 12 beats

The creation of this graph isn’t needed for the story generation and the image generation. It’s just there for this specific tutorial, to highlight the type of story generated by the LLM.

Now, 12 steps isn’t enough to provide a full storyboard.

Let’s ask GPT to split each step in 2 to 4 beats:

I'm looking to break each of the 12 storyboard scenes into a certain number of smaller beats (for example, 2–4 mini-panels each)

ChatGPT’s answer:

You’ll want to give each of your 12 key scenes enough “micro-beats” to carry the pace and emotional weight—in practice that’s 2–4 mini-panels per scene, for a total of:

Minimum: 12 scenes × 2 beats = 24 substeps
Maximum: 12 scenes × 4 beats = 48 substeps
Sweet-spot (avg 3 beats): 12 × 3 = 36 substeps

So now we will have each 12 steps split up into 3 beats.

Scene #	Stage	# Beats	Why?
1	Ordinary World	3	Intro beats: establish workshop → Mira’s routine → hint at longing.
2	Call to Adventure	3	Arrival → plea → Mira’s first flash of resolve.
3	Refusal	3	Shock → fear → almost-action-then-pull-back.
4	Meeting the Mentor	3	Teach → lore reveal → mentor’s challenge stated.
5	Crossing the Threshold	3	Threshold door → Mira steps through → immediate trial tease.
6	Tests, Allies, Enemies	4	Sand wraith fight → fox ally → labyrinth → map clue.
7	Approach to Inmost Cave	3	Forest entry → whispers of doubt → Mira steels herself.
8	Ordeal	2	Shadow-Mira appears → emotional clash → shard seized.
9	Reward	2	Dawn light → shard held high.
10	The Road Back	3	Golems awaken → race uphill → near-miss escape.
11	Resurrection	4	Summit reveal → dragon confrontation → sacrifice → triumph.
12	Return with the Elixir	3	Workshop dawn → maps reborn → world reacts.

The story won’t win any award for originality or even good creativity, but it’s good enough for this tutorial.

We can now ask ChatGPT to convert those 36 beats to a storyboard sheet.

Each description will be used to generate a starting frame.

Now, create a full breakdown of those 36 substeps as a markdown file, with a list of prompts, to generate them with a prompt to video system, following the format in the provided screenshot

Here’s a partial screenshot for the created file:

Partial screenshot of markdown file with substeps

2. Moodboard creation

We need to ask the system to create a moodboard that will be used as a reference for the graphical style. Each time we will ask for a new image, we will also provide the moodboard with the prompt, to steer the image generation in the right direction.

This will help us maintain a coherent style.

Do a few tries until you get a graphical style matching your intent.

You can also provide outside images to GPT at this point if you have something very specific in mind.

Now, for the step 1, create an image that is a moodboard/reference that can be given as a reference for future generation with other LLM

First moodboard attempt

Make the drawing style less complex to read

Second moodboard attempt, less complex

Make it more complex, but less than before, use more diverse color palette

Third moodboard attempt, more complex with diverse colors

Add a little bit more diversity to the colors

Final moodboard style

We will go with that style! Save the image as “moodboard.jpg”

3. Image Generation

Now, the image-by-image generation can start!

We will re-use the same pattern all the time:

Create the starting frame for the " 1. **Substep 1** - **Subject Introduction:** A young cartographer’s apprentice hunched over parchment - **Primary Action:** Sketches the outline of distant stars - **Camera Instructions:** Static overhead shot focusing on the table - **Lighting & Mood:** Warm lantern glow casting soft highlights - **Environment & Props:** Dusty wooden workshop with scrolls, ink pots, quills - **Artistic Style or Lens:** Shot with a 50 mm prime lens for natural perspective ", using the provided image as a reference for the style Moodboard reference for prompt

The idea is to have a systematic prompt :

Create the starting frame for the <insert the description of the frame here>, using the provided image as a reference for the style. Don’t forget to drag and drop the moodboard into ChatGPT’s prompt window in the webpage!

Result of the generation:

First generated frame

(ChatGPT is adding some of its comments to the generated image):

ChatGPT comments on generated image

Now, we can iterate, and generate the second image:

Create the starting frame for the " - **Subject Introduction:** Close-up on the apprentice’s hand holding a quill - **Primary Action:** Traces intricate grid lines on the star map - **Camera Instructions:** Tight close-up on hands and parchment - **Lighting & Mood:** Soft amber lamplight with deep shadows - **Environment & Props:** Wooden desk scattered with compass and ruler - **Artistic Style or Lens:** Shot with an 85 mm prime for shallow depth-of-field " make it consistent with the provided previous frame and the provided moodboard Previous frame reference

Here, I’m providing the previous frame with the moodboard, to make the environment coherent. Since the scene is still inside the same house, I want to keep the same objects, colors and placement. Don’t use the previous frame if there is a cut between the scenes (like for example if the previous scene is in the snow, and the next one in the sand).

Result of the previous generation:

Second generated frame, consistent with first

As you can see, it’s pretty consistent.

3.1 Character Sheet

Consistency for characters is even more important than consistency for the environments.

To enforce that, we will use a tool from the animation industry: the character sheet. It’s a reference sheet, a turnaround of a character, to see it from all angles. Sometimes you have insets with zoom-in on the face, the hands, or other details.

Some examples:

Sintel character sheet example

https://commons.wikimedia.org/wiki/File:Character_Sintel-model-sheet.png

Character design model sheet example

https://upload.wikimedia.org/wikipedia/commons/1/18/Fig._Character_Design_Model_Sheet_C_%28cc%29_2017_03.jpg

1D model sheet example

https://commons.wikimedia.org/wiki/File:1D_model-sheet.png

Now, we will create a character sheet for the various characters in our story:

Create a concept art graphical character sheet for the apprentice Apprentice reference from frame 1

The result:

Character sheet for the apprentice

Of course, create the various character sheets at the beginning of your project, it will help you get consistency from the start!

For the wizard:

Create a concept art graphical character sheet for the wizened traveler Wizard reference image

The result:

Character sheet for the wizened traveler

To enforce image ratio, you must add them into the prompt.

Indeed, during the generation, if you let ChatGPT on its own, it will generate images with different ratios.

Substep 02 image with specific ratio Substep 05 image with square ratio Substep 13 image with portrait ratio Substep 34 image with landscape ratio

3.2. Image generation (with ratio)

Example:

Create the starting frame for the " - **Subject Introduction:** A young cartographer’s apprentice hunched over parchment - **Primary Action:** Sketches the outline of distant stars - **Camera Instructions:** Static overhead shot focusing on the table - **Lighting & Mood:** Warm lantern glow casting soft highlights - **Environment & Props:** Dusty wooden workshop with scrolls, ink pots, quills - **Artistic Style or Lens:** Shot with a 50 mm prime lens for natural perspective " make it consistent with the provided moodboard Moodboard reference

Example with image ratio specified:

Create the starting frame for the " - **Subject Introduction:** A young cartographer’s apprentice hunched over parchment - **Primary Action:** Sketches the outline of distant stars - **Camera Instructions:** Static overhead shot focusing on the table - **Lighting & Mood:** Warm lantern glow casting soft highlights - **Environment & Props:** Dusty wooden workshop with scrolls, ink pots, quills - **Artistic Style or Lens:** Shot with a 50 mm prime lens for natural perspective - **Image ratio:** 3/2 image (1536 by 1024)" make it consistent with the provided moodboard Moodboard reference

Image generated with 3/2 ratio

The final results!

Collage of final generated images - set 1

Collage of final generated images - set 2