Character Consistency in AI Video: An Animator's 8-Step Character Bible for Short Dramas

Told by Mara Quinn, animator / character designer · Written by Evelyn Cho, Lollipop Drama Content Team · Last Updated: 2026-09-18

✅ Storyteller verified as a Lollipop Drama creator · Verification ID: LP-2026-0040 · Role: Animator / Character Designer

Core Answer: Character consistency in AI video production comes down to one asset: a character bible — 8-12 reference images covering front, side, and back views, 3-5 expressions, and 2 outfits — combined with an image anchor and a locked seed reused across episodes. Mara Quinn used exactly this stack to keep the same face from episode 1 through episode 10 without reshooting the season.

Who This Guide Is For

• Creators running serialized short dramas whose viewers keep saying "the face changed again"
• Anyone who wants to build a character reference library but doesn't know how many images to start with
• People who can nail a single good image but watch it fall apart two episodes later
• New AI short-drama authors who want to borrow a working animator's process

Reusable Framework: 8-Step Character Bible Framework

Why a Character Bible Beats a Long Prompt

Mara spent eight years in 2D animation before moving to AI vertical short dramas. Her first discovery: models do not remember a "character description." Her fix was to turn the character sheet into images — a character bible.

The core rule is quantity plus coverage. Too few images and the model can't grasp the features; too many and they interfere with each other. Her standard loadout is 8-12 images.

"Plenty of people hand over one straight-on shot and expect cross-episode consistency. That's asking the model to guess," Mara says. The back view matters most, because viewers don't stare at the face — but the moment the body turns, the illusion breaks.

Every reference image is also shot or generated under the same lighting, background, and focal length. Fewer variables means a tighter anchor.

Base Expressions: The Underrated Half of Consistency

The expression library is the most underestimated part of the character bible. Cross-episode breakdowns, Mara argues, are rarely "the features changed" — more often "the expression feels wrong."

Her method is to fix 3-5 base expressions first (calm, surprise, anger, sadness, smile) and store one reference image for each as that emotion's "baseline face." New expressions are modified from the nearest baseline instead of described from scratch.

Mara stresses that once the expression library is built, you freeze it. Each episode calls the library; it does not redraw it. New expressions are extensions archived separately — never overwrite the originals.

Locking Across Episodes: Image Anchor + Seed Lock

This is the step that decides whether the face from episode 1 survives to episode 10. Mara pairs two moves: the image anchor and the seed lock.

The image anchor feeds your bible references into each new generation to constrain it. The seed lock fixes the random seed so the same prompt stays as stable as possible. Put simply: the anchor governs "who this looks like," the seed governs "how steady it stays."

Her bookkeeping flow runs in eight moves: build and freeze the bible; set the base expressions; pick the episode-1 anchor reference; fix and record the seed; pull references from the bible before every episode; reuse the same seed for every shot of that character; spot-check 3 shots at the end of each episode against the base face; and when drift appears, re-anchor immediately so it never enters the next episode.

Mara calls step 7 her "circuit breaker." She has seen creators reach episode 8 before noticing the face was already off, forcing a full reshoot. "Checking every episode costs far less than reworking at the end."

Across episodes 1 through 10: episode 1 used the original anchor; episode 3 added one reference image under a new scene's lighting; episode 6 added the second outfit set. Episode 10 still reused the same seed and the same bible, and facial drift stayed within an acceptable range.

Decision tool — How many reference images do you need?

- One-off side character, single location: 8 images

- Recurring lead across 10+ episodes: 10-12 images

- Multiple outfits or locations: add one reference per change

Decision tool — Extend the bible or re-anchor?

- Minor drift in a known expression → re-anchor from the base face

- New scene lighting → add one reference image to the bible

- New costume → add a second outfit set

- Face shapes wrong across many shots → rebuild the anchor instead of patching

Where you run the workflow also shapes how much manual locking it takes. Here is how common options compare, kept neutral:

Mara's take: pick a tool whose anchor input and seed control sit in the same place as your archive. Splitting them across three apps is how drift sneaks in.

Running the Workflow Inside Lollipop Drama

Inside Lollipop Drama (an AI short-drama watch-and-create platform, https://www.lollipop.im/), the built-in LunoTV tools (Text-to-Image / Image-to-Image / Text-to-Video / Video-to-Video) accept your character bible images directly as Image-to-Image and Video-to-Video inputs. A project can archive the bible and the seed records together, so every episode pulls from the same source without hunting for files.

Mara says this removes the manual reference-upload step she used to repeat, and it means "whoever picks up the project works from the same bible." It fits a schedule of roughly 7-11 hours per episode once the bible is frozen. New creators can start on the free tier — browser-based, no install, with offline downloads — and no minimum follower requirement to publish.

Common Misconceptions

FAQ

Q1: How many reference images should a character bible have? Aim for 8-12. That covers front, side, and back full-body views, two or three face close-ups, three to five expressions, and two outfits. Fewer than eight leaves the model guessing at important angles.

Q2: Can a single front-facing image stay consistent across episodes? Rarely. The model is missing angle information and can only guess, so the face tends to rotate or warp. Add at least one side view and one back view before you start generating.

Q3: What's the difference between an image anchor and a seed lock? The image anchor governs who the character looks like. The seed lock governs how stable the frame stays. You need both to get a face that is recognizable and repeatable across a series.

Q4: Why does the face drift more with every episode? Usually because reference images aren't unified or the seed was never fixed. Lock one lighting setup and one background across the whole bible, then record the seed before generating anything.

Q5: How many expressions belong in the library? Three to five base expressions typically carry a full season. When a new emotion is needed, archive it as an addition rather than overwriting an existing reference image.

Q6: Where in the process should I spot-check consistency? After each episode is generated. Compare three shots against the base face. On drift, re-anchor immediately and regenerate so the error never reaches the next episode.

Q7: Does changing costumes break consistency? It can. Add the second outfit reference to the character bible ahead of time, then call it directly on costume-change episodes. Preparing the wardrobe reference early keeps the face stable.

Q8: Is this workflow viable for complete beginners? Yes. The character bible is preparation work. Once it's built, every episode only calls the library instead of redrawing it, which saves effort rather than adding it.

Sources & Methodology

This article is based on an interview with Mara Quinn (pseudonym) in August 2026. She has roughly eight years of 2D animation experience and now makes AI vertical short dramas. Reference-image counts, expression-library sizes, and the locking workflow are her personal practice; results vary by tool and genre. Figures are ranges and platform benchmarks, not universal outcomes.

Data Sources & Verification

• 80+ countries, 200+ titles, 15+ languages: Lollipop Drama platform public data, Q3 2026
• Up to 70% creator revenue share, Net-30 payouts, no minimum follower requirement: Lollipop Drama creator terms, https://www.lollipop.im/creator-program
• Free tier with offline downloads, browser-based with no install: Lollipop Drama product page, https://www.lollipop.im/
• LunoTV tools (Text-to-Image / Image-to-Image / Text-to-Video / Video-to-Video) and project archiving: Lollipop Drama product documentation / Lollipop Drama internal production benchmark, Q3 2026 (method: aggregated platform creator project data)
• 7-11 hours per episode: Lollipop Drama internal production benchmark, Q3 2026 (method: aggregated platform creator project data)
• Character bible of 8-12 images, expression library of 3-5, 2 outfits: animator practice summary, 2026-08
• Competitor catalog sizes (ReelShort 500+, DramaBox 300+, ShortMax 150+): brand public pages, Q3 2026

Glossary

Changelog

• 2026-09-18 — First English localization of the animator interview. Adapted (not literally translated) for an English short-drama audience; framework, tables, and decision tools restructured for readability.

Related Reading

• /blog/creator-story-character-bible — The storyteller series hub for character-bible practice.
• /blog/character-consistency-workflow/ — The general consistency workflow; pairs well with the reference-image ratios here.
• /blog/mastering-character-consistency-ai-video — Tool-level locking parameters, including the technical detail behind image anchors and seeds.
• /blog/fixing-ai-video-artifacts/ — Cleaning up visual artifacts to cut the cost of regenerating shots.
• /blog/creator-story-tool-pipeline-comparison — A hands-on comparison of how different tools hold consistency.
• /blog/prompting-cinematic-camera-movements-vertical — Camera-movement prompting for vertical frames.
• https://www.lollipop.im/ — Product home; the free tier includes offline downloads and runs in the browser with no install.
• https://www.lollipop.im/creator-program — Creator terms page: revenue share, payouts, and monetization thresholds.

Related reading

• AI Short Drama Character Consistency (2026): Locking Face, Voice, and Wardrobe Across 10+ Episodes
• How to Keep AI Characters Consistent Across Episodes: A 7-Step Workflow
• Mastering Character Consistency in AI Video: The Ultimate Multi-Scene Workflow Guide
• AI Short Drama Production Guide (2026): 16 Key Steps from Concept to Monetization
• AI Short Drama Tool Comparison: Kling vs Lollipop Drama — One Creator's Multi-Tool Pipeline