Glasses, Hair and Wardrobe Continuity in AI Drama: Lessons From 244 Storyboard Frames
Credits and specs follow TWno1's live billing settings. Last checked: 2026-10-05
In September 2026 we audited 244 AI storyboard frames: glasses were wrong in 58, hair in 130, wardrobe in 106.
The frames came from a 16-episode demo drama, reviewed on 2026-09-29. As shares, that is 23.8%, 53.3% and 43.4%. Other categories were bigger. Actions that ignored the script hit 220 frames, missing or wrong props 186. But look errors are among the ones most often pointed out. Gold-rimmed glasses on, off, on again across three shots. Nobody needs film school to notice that.
This piece takes one glasses failure apart, then covers hair and wardrobe, and ends with a per-frame sheet you can copy.
One failure, four causes
The script said the lead puts on gold-rimmed glasses in frame 7. The storyboard showed her wearing them from frame 3. In frame 7 she put them on again. Some frames that needed glasses had none.
When we traced it back, there was no single bug. Four things were true at once:
| # | Cause | Stage | Fix |
|---|---|---|---|
| 1 | The glasses were written into the outfit description, so every frame inherited them | Character setup | Move glasses out of the outfit text into a per-frame state |
| 2 | Nothing remembered whether she wore them in a given frame | Shot planning | Derive the state from the script and pass it down |
| 3 | Any mention of "glasses" drew glasses, even "not wearing glasses" | Prompt | Delete every glasses word in frames without them |
| 4 | The layout sketch already had glasses, and the render traced it | Sketch | Strip glasses and hair clauses from sketch prompts too |
Leave any one of these in place and the glasses keep flickering. That is also why re-rolling did nothing. A new seed can clear a random glitch. These four fire every time.
Name it and you get it
Cause 3 trips up the most people, because the usual advice points the other way.
The usual advice: if you don't want something, say so. Write "she is not wearing glasses," or put "glasses" in a negative field. In our frames, once the word sat in the description, the model leaned toward drawing it, whatever verb came before it.
Runway says something similar in its Gen-4 video prompting guide (checked 2026-10-05). Negative phrasing is not supported there and can produce unpredictable or opposite results. Runway's example: write "locked camera," not "no camera movement."
The rule has an edge, though. BytePlus's Seedance 2.0 prompt guide (checked 2026-10-05) recommends constraint lines such as no subtitles, no logo, no watermark. Our reading of the difference: subtitles and watermarks are global conditions on the whole clip. Glasses are an object on a character. Name the object and it becomes something to draw.
So:
- Frames with glasses: write "wearing gold-rimmed glasses."
- Frames without: write nothing about glasses. No "without," no "after taking them off."
- Apply the same filter to sketch and reference prompts.
Hair: 130 wrong frames from one "helpful" rule
Hair was the worst of the three at 53.3%. The classic case: the script asks for a low bun, every frame shows long loose hair.
Three causes. A rule in the workflow we used on our demo drama allowed one hairstyle per character for the whole film, and it overrode the script. A word list missed a phrasing, so one way of writing "low bun" was never recognized as a hair clause. And the look reference picked for the character showed loose hair.
The last one is easy to miss. When the text says bun and the image says loose, the image usually wins.
Here is where common advice is half right. "Lock one look per character for consistency" sounds sensible. Continuity, though, means each frame matches the script at that moment. If she leaves home with her hair down and wears it up at the gala, that is the story. Locking one style breaks it.
What we do now:
- Decide hair per frame from the script. No film-wide default.
- Write hair and accessories in one place only, the appearance text. The outfit field holds clothes and nothing else. BytePlus's guide also warns against giving one subject contradictory traits, and two places for one fact will drift apart.
- Use one term per hairstyle for the whole script. Pick "low bun" and never alternate with three synonyms.
- Match the reference image to the text. If the text says bun, the reference shows a bun.
Wardrobe: the planner never read the story
Example: the script puts the lead in a wine-red dress for a dinner. The frame showed a daytime office shirt dress.
In that workflow, the step that planned outfits had received a one-line reason and the character's job. It never saw the script. When its keyword match failed, it fell back to a default outfit, which happened to be office wear.
Two fixes. The wardrobe step reads the original script. And every frame records which outfit it uses, explicitly, instead of guessing from keywords.
Restate the full look on each scene's first line
This rule came from a specific miss. The script said "puts on gold-rimmed glasses" without saying she wasn't wearing any before. The office scene then never restated her low bun, glasses or watch. The frames came out with loose hair and bare eyes.
Two writing rules follow:
- For anything that changes, state the before. If she puts glasses on in frame 7, the script should show she had none earlier.
- On the first line of every new scene, write the whole look again: hair, glasses, accessories, outfit. Even when nothing changed.
Example first line:
[Office, day] She wears her hair in a low bun, gold-rimmed glasses, a silver watch on her left wrist, a gray blazer over a white shirt. She stands at the window.
The per-frame look sheet
Put every rule above into one table. One row per frame, one column per look item that can change, and a column for the script evidence. The rows below are illustrative.
| Frame | Scene | Glasses | Hair | Outfit | Evidence |
|---|---|---|---|---|---|
| 1 | Apartment, morning | None | Loose | Loungewear | First glasses event this episode is "puts on" in frame 7, so none before it |
| 4 | Office, day | None | Low bun | Blazer, white shirt | Look restated on scene's first line |
| 7 | Office, day | Gold-rimmed | Low bun | Same as 4 | Script: "puts on gold-rimmed glasses" |
| 12 | Banquet hall, night | Gold-rimmed | Low bun | Wine-red dress | Script names the dress; no removal event, glasses carry over |
Two inference rules keep the sheet honest. We added both after getting them wrong:
- An event belongs to the person most recently named in the same sentence. In "He watches her put on her glasses," she wears them.
- The first event in an episode sets the state before it. First event "puts on" means none before. First event "takes off" means wearing before.
Each frame's prompt is built from the sheet, not copied from the character profile. A "None" cell means the word never appears.
One side effect to watch. If the character's reference portrait has no glasses but the look calls for them, the two sources fight, and the video can show a faint, see-through pair. Make the portrait after the look is final, from the same appearance text.
Check the frames with specific questions
A correct sheet does not guarantee correct pictures. Before any video is made, go frame by frame and ask concrete questions: Glasses on? Low bun? Wine-red dress? A wrong answer means redraw. Video is animated from the storyboard frame, so the frame caps the clip: a wrong frame usually produces a wrong clip.
Back-facing shots need a different check. No face is visible, so compare hair color and clothing instead.
A TWno1 project stores the character's look and reference images, and later shots reuse that same set. That fixes one cause above, frames pulling from different references. If the reference itself shows loose hair, a low-bun scene needs a low-bun image prepared first.
Limits, and when this is not a good fit
- Text cannot beat a wrong reference image. If the reference shows loose hair, a perfect sheet still gets pulled back. Swap the image first.
- Our numbers describe one drama. 23.8%, 53.3% and 43.4% come from one audit of one 16-episode demo. They are not industry rates. Measure your own.
- Re-rolls don't fix systematic errors. If the same mistake shows up every time, change the script, the prompt or the reference.
- Exact accessory detail is a poor fit. If every shot must show the same frame design up close, AI generation is hard to trust. A real shoot or compositing in post is safer.
- Vendors admit the ceiling too. Google's Veo 3 model card lists full consistency in complex scenes and complex motion as a known limitation.
Point rules and packages are on the pricing page.
Related reading: Face-Similarity Scores for AI Drama Characters: Four Checks and the Thresholds We Use; Vertical Drama Shot Length: Why We Aim for 2.5-4 Seconds and Cap Shots at 6; Keeping One Character Consistent Across a Whole AI Series: References, Prompts and a Per-Shot Look Sheet.
FAQ
Why do a character's glasses keep appearing and disappearing in AI drama?
Glasses flicker in AI drama when several causes stack: glasses written into the outfit text, no per-frame record of whether they are worn, prompts that say "not wearing glasses" (which still draws them), and sketches that already include them. In our 244-frame audit, 58 frames (23.8%) had wrong glasses, and all four causes had to be fixed.
Does writing "no glasses" in a prompt work?
Usually not. Once the word "glasses" appears in a prompt, even inside a negative sentence, the model tends to draw them. Runway's Gen-4 prompting guide also says negative phrasing is unsupported and can give opposite results. For frames without glasses, remove every glasses word instead.
Should an AI drama character keep one hairstyle for the whole series?
Only if the script never changes it. In our 244-frame audit, hair was wrong in 130 frames (53.3%), and one main cause was a rule that forced a single hairstyle per character over the script. Decide hair per frame from the script and write it in one place only.
What columns does a per-frame look sheet need?
A per-frame look sheet needs at least the frame number, the scene, one column for each look item that can change (glasses, hair, outfit, watch or other accessories), and the script line that justifies each state. Build each frame's prompt from the sheet, and leave out any item marked "None."
Why restate the full look at the start of every scene?
Without a restatement, frames tend to fall back to a default look. In one of our scenes the script did not repeat the low bun, glasses and watch, and the office frames came out with loose hair and no glasses. Restate hair, glasses, accessories and outfit on each scene's first line, even when nothing changed.
References
- Runway: Gen-4 Video Prompting Guide (checked 2026-10-05)
- BytePlus ModelArk: Seedance 2.0 series prompt guide (checked 2026-10-05)
- Google DeepMind: Veo 3 Model Card (PDF) (checked 2026-10-05)