How I Keep One AI Anime Character Consistent Across Images, Comics, and Video (With the Exact Prompts)

Nland S — Thu, 15 Oct 2026 08:49:06 +0000

If you've tried to build a series around an AI-generated character, you know the problem. The first image looks great. Then the next one has a slightly different face, the hairstyle drifts, the outfit changes color, and by the time you try a video, she looks like a different person.

In this post I'll walk through the workflow I used to take one original character, Clara, from a character sheet to an anime illustration, a 4-panel manga, and a 6-second video with a spoken line and subtitles. I'll share the exact prompts, the settings, and the small rules that made the biggest difference.

Disclosure: I'm building Comik AI, the tool used in this walkthrough. The techniques themselves (a reusable description block, fixed reference roles, explicit negatives, timed subtitles) apply to most image and video models that accept reference images. This post was drafted with AI assistance.

Here's the 30-second result:

The workflow at a glance

Generate a multi-view character sheet
Save it as a reusable character asset
Use it as a reference for an anime image
Use it again for a 4-panel manga
Use it again, with the image as a scene reference, for a video shot with voice and subtitles

Step 1: Start with a multi-view character sheet

A single portrait isn't enough reference. The model also needs to know what the character looks like from the side and from behind. Otherwise it invents those details differently every time.

Clara's character sheet, made in Anime Character Creator with GPT Image 2.5 Flare.

In Anime Character Creator you can describe the character in text or start from a reference image. You also give the character a name, which becomes the name of the saved asset.

Describing a character in text. The character name becomes the asset name.

Clara herself was created from an uploaded reference image with the built-in Use Default Prompt preset, which asks for a front-facing upper-body portrait plus full-body front, side-profile, and back views. A text description like the one above produces the same kind of sheet.

Step 2: Reuse it as a character asset

The sheet is added to My Assets → Character automatically. From then on, any generator that accepts references can pick it from the reference picker.

Picking Clara from My Assets in the reference picker.

Step 3: Put the character in a new scene

The key trick is a canonical description block: one paragraph that describes the character's appearance, pasted word for word into every prompt. The reference image carries the look; the text block tells the model which details matter.

Settings: Anime Image Generator, GPT Image 2.5 Sunburst, 16:9, Clara as the only reference.

Full image prompt

The girl from the reference image (Clara). Keep her face, hairstyle and outfit exactly the same. Clara: a young woman in her early twenties with long, warm chestnut-brown wavy hair styled half-up with a braided crown at the back, center part and soft face-framing strands; warm brown eyes, soft rosy blush, gentle friendly smile. She wears a light periwinkle-blue knit cardigan with small tan buttons and long sleeves over a white collared shirt with the collar out and an open neckline, a chocolate-brown high-waisted pleated mini skirt, sheer taupe tights and dark brown block-heel loafers.

Scene: a cozy secondhand bookstore on a rainy afternoon. Clara sits sideways on a wooden window seat with her legs crossed, reading an old cloth-bound hardcover book held in both hands, with a soft, absorbed smile. Behind her is a tall window with raindrops and rain streaks on the glass and blurred street lights outside. Warm amber light comes from a brass reading lamp on the left. Wooden shelves are packed with old books, and a small stack of books and a ceramic mug of tea sit on the windowsill.

Framing: full shot, eye level, Clara placed right of center, calm open space on the left third of the frame.

Soft semi-realistic anime illustration, clean fine linework, soft cel shading, warm pastel palette, delicate skin rendering, high detail, cinematic warm lighting, gentle rim light from the window.

Show her only once. No text, no captions, no logos, no extra people, no character-sheet panels or split views.

Two lines did most of the work:

Keep her face, hairstyle and outfit exactly the same. sets the priority.
Show her only once. ... no character-sheet panels or split views. prevents a common failure: when the reference is a multi-view sheet, models like to reproduce the sheet layout instead of drawing a single scene.

Result: the same braided half-up hair, periwinkle cardigan, and pleated skirt, in a new scene.

Step 4: Reuse it in a 4-panel manga

For the comic, I wrote one line per panel and spelled out every piece of text that should appear. Comic text is where models improvise the most, so the prompt also says which text is allowed.

Settings: AI Comic Generator, GPT Image 2.5 Sunburst, 16:9, 4 Panels.

Full manga prompt

A black-and-white Japanese manga page with 4 panels in a 2x2 grid, starring the girl from the reference image (Clara). Keep her face, hairstyle and outfit identical in every panel. Clara: a young woman in her early twenties with long, warm chestnut-brown wavy hair styled half-up with a braided crown at the back, center part and soft face-framing strands; warm brown eyes, soft rosy blush, gentle friendly smile. She wears a light periwinkle-blue knit cardigan with small tan buttons and long sleeves over a white collared shirt with the collar out and an open neckline, a chocolate-brown high-waisted pleated mini skirt, sheer taupe tights and dark brown block-heel loafers.

Panel 1 (top left): wide shot inside a rainy secondhand bookstore. Clara reaches up and pulls an old hardcover book from a tall wooden shelf.
Panel 2 (top right): close-up. She opens the book and a small folded paper note slips out from between the pages, falling.
Panel 3 (bottom left): extreme close-up of her hands unfolding the note. Neat handwriting on the note reads: "Make something today."
Panel 4 (bottom right): medium close-up. Clara holds the note against her chest with a warm, surprised smile. Speech bubble: "Who left this...?"

Style: classic shoujo manga, clean ink lines, screentone shading, expressive eyes, crisp black panel borders with white gutters.

Only the note and the one speech bubble contain text. No color, no extra characters, no sound-effect lettering, no watermark.

Result: all four panels are composed into one comic image.

Step 5: Bring it to life in video, with voice and subtitles

For the video I used Reference to video mode with two images and a voice:

Image 1: the character sheet, for identity
Image 2: the bookstore image from Step 3, for the scene and lighting
Voice: Soft-Spoken Lady from the Voice Library

Each reference is inserted into the prompt as an @ mention, so the prompt can say exactly which reference does what.

Settings: AI Anime Video Generator, Seedance 2.5, Reference to video, 6s, 16:9, 1080p, with audio.

Full video prompt

Character Reference: @Image 1
Scene Reference: @Image 2
Voice Reference: @Soft-Spoken Lady

Create a 6-second cinematic anime reference-to-video shot.

Main Character:
Clara, the girl in @Image 1: a young woman in her early twenties with long, warm chestnut-brown wavy hair styled half-up with a braided crown at the back, center part and soft face-framing strands; warm brown eyes, soft rosy blush, gentle friendly smile. She wears a light periwinkle-blue knit cardigan with small tan buttons and long sleeves over a white collared shirt with the collar out and an open neckline, a chocolate-brown high-waisted pleated mini skirt, sheer taupe tights and dark brown block-heel loafers.

Character Consistency:
Strictly preserve the exact character identity from @Image 1:
- same face identity
- same hairstyle and braided crown
- same facial features
- same outfit and colors
- same body proportions
- same visual style
Do not redesign the character. Use @Image 1 only for her appearance. Never show the character-sheet layout or multiple views.

Location:
The cozy secondhand bookstore from @Image 2 on a rainy afternoon: a wooden window seat with cushions, a tall rain-streaked window with blurred street lights outside, warm amber light from a brass reading lamp on the left, and shelves of old books behind her. Raindrops slide slowly down the glass.

Visual Style:
High-quality 2D anime film look, soft semi-realistic anime rendering, clean fine linework, soft cel shading, warm pastel palette, painterly background, soft depth of field, matching @Image 1 and @Image 2.

Camera:
Virtual anime camera, 16:9. Gentle, steady, elegant movement. Keep Clara's face and the note in the upper two-thirds of the frame at all times, leaving the lower third clear for the subtitle. No 3D CG look. No live-action look. No shaky or sudden camera movement.

Timeline:
00:00-00:02
Medium close-up, eye level. Clara sits on the window seat holding a small unfolded paper note in both hands, eyes down reading it, a smile slowly forming. Rain slides down the window behind her. No text on screen.
00:02-00:04
She slowly lifts her eyes from the note and looks straight into the camera with a warm, knowing smile. The camera pushes in slowly to a close-up. No text on screen.
00:04-00:06
Close-up. She raises the note beside her cheek, tilts her head slightly and says softly: "Was this you?" Then she gives a small, light laugh. The camera holds still.

Dialogue (English, spoken once, at 00:04):
"Was this you?"

Animation:
- realistic lip movement synchronized with the dialogue
- natural blinking
- subtle eye movement from the note to the camera
- slight head tilt
- gentle hair and cardigan motion
- smooth, simple hand movement raising the note

Audio:
Steady soft rain on the window glass, the faint rustle of the paper note, muted distant street ambience. Clara's voice: soft, warm, young female voice, natural English. No background music. No narration. No other dialogue.

Subtitle Requirements:
Subtitle text, exactly, as one line: Was this you?
Spell it exactly as written: capital W, lowercase letters, one question mark, no quotation marks, no extra words.

Subtitle Timing:
- No subtitle and no text of any kind from 00:00 to 00:04.
- The subtitle appears exactly when she starts speaking at 00:04.
- It disappears right after she finishes the line, before her laugh.

Subtitle Position:
- lower center of the frame
- slightly above the bottom edge, with a safe margin
- centered horizontally

Subtitle Style:
- clean cinematic movie subtitle
- white text with a thin black outline or subtle shadow
- simple modern sans-serif font
- easy to read

Subtitle Restrictions:
- Do NOT create speech bubbles.
- Do NOT create manga text effects.
- Do NOT add any other captions, logos, labels or watermarks.
- The subtitle must look like a real cinematic video subtitle.

Goal:
A warm, intimate anime moment, as if Clara just discovered that the viewer left her the note: gentle, charming and cinematic.

Two parts of this prompt are worth copying:

A timeline with at most three beats. 0–2s reading, 2–4s looking up at the camera, 4–6s the spoken line. Fewer, clearer beats make the motion more predictable.
Locked subtitle timing. I wanted to reuse the first seconds as the opening shot of a promo, so the prompt says that no text may appear before 00:04, and it spells out the subtitle exactly ("capital W, one question mark"). Burned-in subtitles can't be removed later, so a short line is safer than a long one.

Result: Seedance 2.5, 6 seconds, 1080p, with audio.

Checking the clip with ffmpeg

Before using the clip, I checked two things: no text appears before the dialogue, and the subtitle lines up with the speech.

# A 4x3 contact sheet: one frame every 0.5 s
ffmpeg -i clara.mp4 -vf "fps=2,scale=480:-1,tile=4x3" -frames:v 1 contact.png

# Where does the speech start and end?
ffmpeg -i clara.mp4 -af "silencedetect=noise=-30dB:d=0.2" -f null - 2>&1 | grep silence_

In this clip, the speech runs from about 3.8s to 4.8s and the subtitle is visible from 4.0s to 5.0s, which is close enough. The first 3.5 seconds contain no text, so they work as an opening shot.

What made the difference

Character sheet first. Front, side, and back views give the model something to check against when the camera angle changes.
One description block, pasted everywhere. The same words go into the image, comic, and video prompts.
Say what not to do. "No character-sheet panels" fixed the most common failure.
Give references fixed roles. Image 1 is always the character; Image 2 is the scene.
Spell out every visible word. Speech bubbles, notes, and subtitles.

One caveat: references improve consistency, but they don't guarantee identical results every time. When something drifts, adjust the prompt and regenerate.

Try it

If you want to try this workflow with your own character, you can start here: Create a character in Comik AI.

How do you handle character consistency across models? I'd love to hear what works for you.

DEV Community: Nland S