Reasoning Image Models for Storyboards and Consistent Characters

Storyboards are where image generators used to fall apart. One frame is easy. Twelve frames with the same character, the same jacket, the same room and a camera that moves in a way that makes sense is a different problem. Early text-to-image models treated each prompt as a fresh roll of the dice, so the hero's face changed every panel and "she stands on the left, he enters from the right" came back mirrored half the time.

A newer class of image models handles this better because it does more than match words to pixels. This article explains what changes with reasoning image models, where they still slip, and a workflow for getting consistent characters across a full board.

Who needs a storyboard anyway

Anyone who plans images in sequence: a director pitching a scene, an agency mapping a 30-second spot, a comic artist roughing out pages, or a creator planning a recurring character for short videos. Not every video needs a hand-built board, though. A faceless channel can hand the whole job to an AI faceless reels generator that writes the script and builds the voiceover, captions, visuals, music and edit from a topic. Storyboarding pays off when you want to direct each frame yourself, and above all when a character has to be recognisable from one frame to the next.

What "reasoning" means for an image model

Several recent image models are built on, or work alongside, a multimodal language model. Instead of mapping a prompt straight to an image, they interpret the request first: what objects are in the scene, how many, where they sit, who is doing what to whom, and what the reference images contribute. Then they render.

For storyboards, that shows up in four practical ways:

• Spatial instructions stick. "Wide shot, she is in the left third, the door is behind her on the right" is followed far more often.
• Counts and props hold. Two cups on the table stay two cups when you ask for the next frame.
• References are understood, not pasted. A character sheet tells the model who the person is, so it can show them from a new angle instead of copying the reference pose.
• Edits are targeted. "Same frame, change only her expression to surprise" changes the expression and leaves the lighting, outfit and background alone.

uni-1 is built around this kind of work: a multimodal reasoning image generator for prompting, reference-guided generation, editing and character consistency, used as one loop of prompt, reference, edit and refine rather than one-shot generation.

A workflow for consistent characters across a board

1. Build the character sheet first

Before any story frame, generate the character on a plain background: front view, three-quarter view, side view, and a few expressions. Fix the details that tend to drift: hairline, eye colour, one distinctive accessory, outfit colours.

Approve the sheet before moving on. Every later frame will use it as a reference, so any flaw here repeats twelve times.

2. Write the board as a shot list

For each frame, write four lines:

• Beat: what happens in the story.
• Framing: wide, medium or close-up, and the camera angle.
• Action: what each character is doing, in plain verbs.
• Continuity: what must match the previous frame (props, time of day, which hand holds the bag).

The continuity line is the one most people skip and the one that saves the most rework.

3. Generate frame by frame with references

Attach the character sheet to every generation. For frames in the same location, also attach the previous approved frame as a setting reference, and say what it is for: "Use image 1 for the character, image 2 for the room and lighting."

Keep one character description in text as well, and paste it word for word into every prompt. Text and reference together are more stable than either alone.

4. Edit instead of rerolling

When a frame is nearly right, do not regenerate it from scratch. Ask for a targeted edit: move the camera slightly lower, close her hand around the cup, turn his head toward the window. A fresh roll changes ten things to fix one. An edit changes one.

5. Re-anchor to the sheet, not the last frame

A common trap: using frame 7 as the reference for frame 8, frame 8 for frame 9, and so on. Small changes add up, and by frame 12 the character has drifted. Always reference the original character sheet for identity, and use neighbouring frames only for setting and lighting.

6. Run a continuity pass

Lay the frames out in order and check:

• Screen direction: if she exits right, she enters the next shot from the left.
• The imaginary line between two characters in a conversation is not crossed without a reason.
• Props are in the same state (the cup is still half full, the door is still open).
• Time of day and light direction stay consistent within a scene.

Not every frame needs a character

Some beats in a board carry information rather than action: a statistic, a date, a key phrase. Forcing a character into those frames often makes them weaker. Many explainer and faceless videos cover these beats with motion graphics instead, and tools are appearing for exactly that. AIBroll, for example, is building a generator that reads a script and turns its lines into kinetic type, animated charts and icon motion; it currently has a waitlist. In your board, mark those beats as graphics so you do not spend generations on them.

Where reasoning models still slip

Better reasoning does not make every problem go away. Plan around these:

• Hands and small props are still the most common errors. Check them in every frame.
• Text inside the image is far better than it was, but long signs or labels still go wrong. Keep in-frame text short or add it later.
• Crowds of named characters. Two or three recurring characters are manageable. Five in one frame invite swapped faces and outfits.
• Long sessions. If the model's output starts to drift after many rounds of edits, go back to the sheet and regenerate the frame cleanly.

From board to motion

Approved frames rarely stay still. They become first frames for image-to-video, pitch animatics or the scene images in an animated short. If you are new to that step, it helps to see how other creators phrase a still scene as a short moving shot. The ToonBee gallery collects AI cartoon video prompts sorted by category and model, and any scene there can be reused as the starting point for your own clip, so it doubles as a reference when you write motion prompts for your frames.

Whatever tool you use for motion, keep the same character description and references you used on the board, and the consistency carries through.

Getting started

Start small: one character sheet, a six-frame board, and one continuity pass. Once that loop feels reliable, scale up to longer sequences. uni-1 works on credits rather than a subscription, so you can match a package to how often you generate, revise and ship. See the options on the pricing page.