Keeping One Character Consistent Across a Whole AI Series: References, Prompts and a Per-Shot Look Sheet

TWno1 Editorial · Published 2026-10-05

Credits and specs follow TWno1's live billing settings. Last checked: 2026-10-05

Hair is the detail AI drama gets wrong most: in 244 storyboard frames from a 16-episode series, 130 had the wrong hairstyle (September 2026 audit).

That is 53.3%. Outfits were wrong in 106 frames and glasses in 58. Everyone knows to watch the face. What wrecks a series more often is everything around it: glasses that show up three shots early, hair that should be pinned up but falls loose, an evening dress that turns back into office wear. Here are four habits that keep one character intact across a series. Three come from vendor documentation. The fourth we added after the audit.

The audit, in numbers

This was a full storyboard audit of a 16-episode demo series, dated 2026-09-29. The three look-related categories:

The faces had their own problem. In the same series, the two male leads scored 0.537 on face similarity and the two female leads 0.544, which made them hard to tell apart. Our working rule since then is that any two characters must score below 0.40. That number came from one open-source face-recognition model and only holds for that model. Change the model and you need a new threshold.

Habit 1: build the character once, with the right assets

Each character should have one master record for the whole series: a short appearance text and a small set of reference images. Every shot pulls from it. Do not rewrite the description episode by episode.

BytePlus's Seedance 2.0 prompt guide is specific about what goes into that record:

• A separate face close-up as the face reference, plus one full-body outfit image.
• A name for the subject in the prompt, so the text and image line up.
• The most important asset first.
• No multi-angle sheets (front, side and back on one canvas). BytePlus says these can make the same character appear twice in the frame, and that this cannot be fully prevented.

That last point runs against a lot of standard advice. Turnaround sheets are a staple of hand-drawn animation. In Seedance's reference slots they can produce a twin. Keep the turnaround for your team. Feed the model the face close-up and the full-body image.

Settle the art style at this stage too. BytePlus warns that realistic reference images pull a requested 2D or 3D animated style toward live action unless the prompt names the style or you convert the reference first.

Habit 2: attach references to every shot, but do not max them out

How many reference images each model takes varies a lot (docs checked 2026-10-05):

The cap is not the goal. BytePlus recommends about four or five assets per request: one or two character images, one scene, one camera-motion clip, one audio clip. It also warns that more than four reference people lowers stability. A scene with three characters, each with one face close-up, plus a scene image already sits at the recommended count.

Kling handles one part of this better. Its Elements feature binds 2 to 4 images of one subject into a reusable unit and can attach a 5- to 30-second voice. The Seedance docs we read have no equivalent subject feature. You attach images in order every time and refer to them as "Image 1," "Image 2" in the prompt.

Real photos are out. The Seedance 2.0 series and 2.5 both refuse reference images or videos containing real human faces. BytePlus points to authorized portrait assets or a digital-character library instead.

Habit 3: the prompt carries the action, the references carry the look

Runway's Gen-4 guide says it plainly. Describe the motion, not what the input image already shows, and call the subject "the subject" or use a pronoun. The reason is simple. If the image shows short black hair and the text says "short black hair," the best case is redundancy. The worst case is a small mismatch the model tries to satisfy both ways.

There is an exception the usual advice skips. Kling's prompt guide recommends the opposite for multi-character dialogue: pair each line with a description of the speaker (clothing, hair, role), or the model loses track of who is talking.

Our compromise:

• One person in frame: name and action only. The references handle the look.
• Several people in frame: give each a minimal tag that tells them apart ("Mara in the charcoal blazer"). One distinguishing trait, nothing more.
• Any shot: anything that changes during the story (glasses, hairstyle, jacket on or off) stays out of the master record and goes into the per-shot sheet below.

Habit 4: a per-shot look sheet

The first three habits come from vendors. This one we were forced into by the audit.

Take the 58 glasses errors. Four causes stacked up:

1. The appearance text had glasses baked into the outfit description, so every frame wore them.
2. The workflow we used on our demo drama kept no record of whether glasses were on in a given frame.
3. Name it and it gets drawn. Any mention of "glasses" in a prompt, even "not wearing glasses," tended to put glasses on the face. Negative phrasing did not stop it.
4. The rough storyboard sketch already had glasses, and later steps followed it.

Hair went wrong the same way. Our setup at the time allowed each character one hairstyle for the whole series, which overrode a script that called for a low bun. The same style was also written several ways, and some versions were not recognized. The 106 outfit errors came from choosing outfits off a one-line job description without going back to the script. When nothing matched, the character got a default outfit.

The fix is one idea: look state follows the shot, not the character. In practice it is a table:

The rules behind it:

• If it is off, leave it out. Omitting glasses and writing "no glasses" are different. Only omission works.
• Hair and accessories live in one field. The outfit field holds clothes only, so two fields cannot contradict each other. Use one fixed term per style. Pick "low bun" and never switch to "hair pinned low."
• Write the state before the change. If shot 7 says "puts on gold-rim glasses," shots 1 to 6 are glasses-off. Our rule: the first event in an episode sets the state before it.
• Restate the look at the start of each scene. Hair, glasses and outfit, in full, on the first line. We skipped this once in an office scene and got loose hair and no glasses.
• Check the output. The sheet only guarantees the instructions are right. Look at full-size frames. A thumbnail contact sheet misses details.

Before you trust any tool's reference slot, run the dumbest test there is. Swap only the outfit reference, keep the prompt and settings fixed, and see whether the clothes change. If they don't, that image is doing nothing, and attaching three more won't help.

Limits: when this approach is not a good fit

• It does not fix face drift. The look sheet handles styling. Faces drifting during a head turn, a hug or a hand across the face is a model limit. Google's Veo 3 model card lists keeping full consistency in complex scenes and complex motion as a known limitation, and a 2026 preprint says commercial video models still suffer identity drift. That takes references plus review and re-rolls.
• Overkill for a one-off clip. A single short with three to five shots can be checked by eye. Skip the sheet.
• Long takes raise the risk. A May 2026 preprint links long-video drift to duplicated characters and lost attributes. If the look must hold, keep shots short.
• Adult fictional characters only. None of this applies to real people or minors. TWno1's content policy bans using a real person's photo, name, likeness or voice for sexual or nude content.

The master record from Habit 1 only works if nobody rewrites it per episode. TWno1 keeps a character's appearance text and reference images in the project, so shot 1 and every later shot draw on the same set. To set up one character's master record, open a project in Studio.

Related reading: Seedance 2.5 vs 2.0: Clip Length, Reference Limits, 4K and Price Compared; Seedance Draft Mode: When a 480p Preview Before 1080p Actually Saves Money.

FAQ

What breaks character consistency in AI dramas more often, faces or styling?

In our 2026-09-29 audit of 244 storyboard frames from a 16-episode AI drama, styling errors were widespread: hair was wrong in 130 frames (53.3%), outfits in 106 (43.4%) and glasses in 58 (23.8%). Face-similarity checks only cover faces, so glasses, hair and outfits need a per-shot look sheet.

How many character reference images should I give Seedance?

BytePlus's Seedance 2.0 guide recommends about four or five assets per request, including one or two character images (a face close-up and a full body), and warns that more than four reference people lowers stability. As of 2026-10-05 the cap is 9 images on Seedance 2.0 and 30 on Seedance 2.5, but the cap is not the recommendation.

Can I use a turnaround sheet as a Seedance character reference?

It is not recommended. The BytePlus Seedance 2.0 prompt guide warns that multi-angle character sheets can make the same character appear twice in one frame, and that this cannot be fully prevented. Use a separate face close-up and a full-body outfit image instead.

If I attach reference images, should the prompt still describe the character's looks?

For a single character, no. Runway's Gen-4 guide recommends describing motion rather than repeating what the input image shows. Multi-character dialogue is the exception: Kling's guide suggests pairing each line with the speaker's clothing or hair, and we add one short distinguishing tag per character.

Why does a character still wear glasses when the prompt says "not wearing glasses"?

In our review logs, any mention of glasses in a prompt, including "not wearing glasses," could make the model draw them, because negative phrasing did not block it. When a character should not wear glasses in a shot, remove every reference to glasses and mark that shot as "omit" on the per-shot look sheet.

References

• BytePlus ModelArk: Seedance 2.0 series prompt guide (checked 2026-10-05)
• BytePlus ModelArk: create a video generation task (API) (checked 2026-10-05)
• Runway: Gen-4 video prompting guide (checked 2026-10-05)
• Kling: Element Library 3.0 user guide (checked 2026-10-05)
• Kling: AI prompt guide (checked 2026-10-05)
• Gemini API: Veo video generation (checked 2026-10-05)
• Google Veo 3 model card (PDF) (checked 2026-10-05)
• arXiv 2605.18733: character drift in long video (preprint) (checked 2026-10-05)
• arXiv 2608.20749: identity drift in commercial video models (preprint) (checked 2026-10-05)

• AI short drama
• Tutorials