Character Sheet: Lock Down Appearance, Clothing, and Personality
This page covers Asset 1: Character Sheet from the asset pack. 2.5 Shot Continuity Check ended with a loose thread: Old Chen in the storyboard is still just one sentence, and the AI redraws his face every time it generates. This page locks down each on-screen character's appearance, clothing, and personality into one fixed description—write it once, and every subsequent prompt references that same paragraph. AI character design doesn't rely on inspiration; it relies on a sheet that only lists "paintable" traits (the sheet template is designed for this course's teaching—just plug in your own character).
What You'll Be Able to Do After This Lesson
- Circle all on-screen characters from the storyboard and list "who needs a character sheet"—the voice-only daughter is explicitly skipped
- Fill out a complete character sheet using the four field blocks of appearance, clothing, carried items, and personality (this course's template)
- Translate unpaintable personality words like "plain and kind" into paintable character description words such as facial features and posture
- Compress the character sheet into a fixed description paragraph of 100 characters or fewer, save it as a separate file, and reference the same paragraph in every shot of the film
Your Current Struggle
- "In the same film, the protagonist in shot 02 and shot 06 doesn't look like the same person at all—even after regenerating five times, it's still two different faces."
- "The prompt says 'a delivery rider,' but the age, build, and vest color all come down to a gacha roll."
- "You try to write a serious persona description, and all you squeeze out is 'middle-aged man, plain, kind'—the result looks completely generic."
When to Use This Technique
When you need to:
- Make the same character look as much like the same person as possible in every shot
- Prepare a stable text basis for the multi-angle reference images in the next lesson
- Let yourself two months from now (or a collaborating friend) know exactly what this character looks like
And you don't want to:
- Re-describe the character for every shot, with a different description each time
- Generate ten shots and roll ten different faces, burning time and credits on retry after retry
- Only discover during editing that the vest changed color or the cake box vanished, ruining entire shots
🎒 Preparation Before You Start
- Completed 3.0 Character and Scene Assets: you know the asset pack delivers three items, and this page delivers Asset 1
- The storyboard is finalized: 2.4 Storyboard—only with a shot list do you know how many on-screen characters need sheets
- Passed the 2.5 Shot Continuity Check: what the character holds in each shot has been sorted out (Old Chen's cake box was exactly what that check caught)
- A blank text file or a sheet of paper—this page is all paperwork, no generation tools needed
💡 When do you actually need to open a tool?
What this page delivers is a pure-text asset. The first real image generation happens in the next lesson, 3.2 Character Reference Images; for basic image-generation operations on the local route, see First Generation: The Default Workflow Explained Node by Node in the comfyui section.
Core Idea
Why you must have a character sheet first: the model doesn't remember the previous shot.
The official production workflow itself is "story conception → storyboard → character references → generate multiple short segments → edit and assemble" (Runway's official workflow; see 1.1 Course Map for details). Each shot is generated separately—the moment a shot is done, the model forgets Old Chen, and the next shot has to get to know him again from the prompt. Even if two prompts are identical, the generated results may still paint two different faces.
So the consistency of AI character design doesn't rely on model memory, but on consistent input:
| Without a character sheet | With a character sheet | |
|---|---|---|
| Prompt for shot 02 | "A delivery rider looking at his phone in the rain" | Fixed description paragraph + "head down looking at phone in the rain" |
| Prompt for shot 04 | "A delivery rider rushing up the steps carrying a box"—doesn't match 02 | The same fixed description paragraph + "rushing up the steps hugging a cake box" |
| Generated result | Two faces, two outfits | The same basis; small drift is still possible, but it's traceable and fixable |
Let's be clear about the boundary: the character sheet improves consistency, but it is not a guarantee of "a face that never changes across shots"—Runway Gen-4 official help also treats reference images as a means of "guiding" the character and recommends iterating step by step. Residual drift in face shape and clothing gets caught by the 2.5 continuity check, with 4.3 first-and-last-frame control as the fallback.
There's also a set of terms that are easy to confuse:
| Term | Who it's for | Example |
|---|---|---|
| Persona | Story background for the audience | "Old Chen is a dad who dotes on his daughter—harsh mouth, soft heart" |
| Character sheet (this page) | For the generation model, and for yourself | "Square face, short black hair, yellow rider vest, white cake box in his arms" |
The persona makes the story memorable; the character sheet makes the model paint the right face—an AI persona description only counts as completed AI character design once it's turned into "paintable" traits.
Follow Along
Below is a full demonstration using Old Chen, the protagonist of "Still Pretty When It Collapses" (this course's teaching example—you can follow along the whole way with your own character).
Step 1: Circle the on-screen characters from the storyboard
Why: The character sheet is only written for characters the audience can see. List them first, then start writing, to avoid two extremes—missing a character who appears on screen, or wasting a sheet on a voice-only character.
How: Open the 2.4 Storyboard, go through the "frame" column shot by shot, and record the shot numbers where each character appears under that character's name:
| Character | Appearances | Write a card or not |
|---|---|---|
| Old Chen (delivery rider) | Full body or close-up face: 02, 04, 06, 08, 10; partial: 05 (shoes), 09 (hands) | ✅ Yes, one card covers the whole film |
| Daughter | Only voice-over throughout, never shows her face (per the header convention in the 2.2 script) | ❌ No, mark as "voice-over only" |
You should see: a character list where every on-screen character is followed by the shot numbers they appear in. Partial shots (the shoes in 05, the hands in 09) must be recorded too—shoes and hands can also break continuity, and they are aligned via the clothing fields on the card.
- Character list complete: on-screen characters each have their shot numbers, and characters who never show their faces are annotated with the reason
Step 2: Fill in the appearance fields—AI character profiles should only contain words that can be drawn
Why: generative models cannot draw "plain" or "weathered"—they only recognize descriptions that translate into pixels. The example prompt in the Wan2.2 official repository README (the white cat on a surfboard) is written exactly that way: a fluffy white cat, wearing sunglasses, sitting on a surfboard, relaxed expression—not a single abstract word, every phrase can be drawn directly (this example is broken down line by line in 4.2 Motion and Camera Prompting).
How: fill in appearance field by field:
| Field | How to fill it for Old Chen (teaching example) |
|---|---|
| Age range | around 40 years old |
| Face shape and skin tone | square face, darkish and rough skin |
| Hairstyle and hair color | short black hair |
| Build | medium, on the thin side |
| Distinctive features | a layer of short stubble on the chin, fine lines at the corners of the eyes |
You should see: every cell can be handed to an artist and drawn directly from it. "Around 40 years old" is better than "middle-aged"—"middle-aged" is too broad a range; the model might draw someone in their 30s or in their 50s.
Step 3: Fill in the clothing and personal items fields, locking them down to color and style
Why: clothing is the most common continuity break—a yellow vest in one shot turns blue in the next, because "delivery rider" only constrains the occupational category and leaves color and style entirely to the model's randomness. The 2.5 Continuity Check once caught this: the visual column of shot 04 omitted the cake box, so the generated Lao Chen would have run empty-handed.
How: fill in clothing item by item, leaving nothing out:
| Field | How to fill it for Old Chen (teaching example) |
|---|---|
| Top | yellow rider vest over a dark blue long-sleeve shirt |
| Pants | black long pants |
| Shoes | worn grayish-white sneakers |
| Personal items | white cake box tied with a red ribbon—never leaves his hands throughout the film |
You should see: every garment has a color and a style. The personal item that "never leaves his hands throughout the film" is marked separately—it goes into the fixed description paragraph; items that appear only in individual shots (like the phone, used only in 02, 03, 04, 10) are left to the shot prompts.
- Clothing locked down to color and style; the never-leaves-his-hands personal item has been marked separately
Step 4: Fill in the personality field, translating it into expressions and posture
Why: personality cannot be drawn, but the default expressions and posture determined by personality can be. Also, the reference images in the next lesson follow the official recommendation of "neutral expression, even lighting" (the reference-image input advice in Runway Gen-4 Help)—so personality must be translated into features that are visible even in a neutral state, not into moments of emotional outburst.
How: write personality words on the left, translate them one by one on the right:
| Personality words (for yourself) | Translated into drawable features (for the model) |
|---|---|
| Dotes on his daughter | gentle eyes and brows, pronounced fine lines at the corners of the eyes |
| Panics when things go wrong, keeps a stiff upper lip | corners of the mouth slightly pressed down; quick little steps when anxious |
| Years of wind and sun exposure | darkish, rough skin (already in the appearance fields, not repeated) |
You should see: the right column is all expressions, posture, and habitual movements—only the right column goes into the prompts later. The left column stays on the card for your own grasp of the character, and it will also come in handy when choosing a voice in 5.1 Voiceover and Licensing.
Step 5: Compress it into a fixed description paragraph, referenced throughout the film
Why: the profile card is a table for you to read, while the generation tool consumes a continuous piece of text. The fixed description paragraph is the condensed form of the card—from now on, the character part of every shot's prompt starts from this paragraph, word for word, unchanged.
How: string the results of steps 2–4 into a paragraph of 100 characters or fewer, and save it to a separate file (e.g., 角色卡_老陈.txt; the asset naming convention will be unified in 3.5 Asset Naming and Version Management):
A Chinese delivery rider around 40 years old, square face, short black hair, darkish rough skin, a layer of short stubble on the chin,
fine lines at the corners of the eyes, gentle eyes and brows, corners of the mouth slightly pressed down; wearing a yellow rider vest, a dark blue long-sleeve shirt, black long pants,
and worn grayish-white sneakers; holding a white cake box tied with a red ribbon in his arms.You should see: this paragraph contains no weather, no actions, no shot numbers—"rainstorm", "rushing up the steps", "looking down at the phone" all belong to the visual descriptions of specific shots, to be added in 4.2 Motion and Camera Prompting. The fixed description paragraph is only responsible for "who this person is, what they look like, what they're wearing, and what they're carrying".
- The fixed description paragraph has been saved to a separate file; paste the whole paragraph when generating reference images in the next lesson
Checkpoint ✅
- One profile card per on-screen character; characters who never show their faces are annotated in the list as "voice-over only, no card"
- The fixed description paragraph contains no undrawable words like "kind" or "weathered"—every line can be handed to an artist to draw from
- Clothing locked down to color and style; the cake box that never leaves his hands throughout the film is written into the fixed description paragraph
- Read the fixed description paragraph through once: no weather, no actions, no shot numbers
- The fixed description paragraph is saved to a separate file—in the next lesson, 3.2 Character Reference Images, just paste the whole paragraph
Pitfall Reminders
❌ Writing "plain and kind personality" directly into the prompt
Symptom: The prompt contains "an honest, kind-hearted middle-aged man," but the generated face has a blurry "random passerby look," and the two generations differ greatly.
Cause: The model cannot draw "kindness"—it can only draw specific eyebrows and eyes, wrinkles, and posture. Abstract words amount to handing the decision over the character's appearance back to random card draws.
Correct approach: Keep personality words in the card for your own reference, and translate them into drawable features such as "gentle eyes, corners of the mouth slightly downturned, holds things against his chest when carrying them," then put those into the fixed description block.
❌ Writing scene states into the character card
Symptom: The fixed description block contains "hair soaked by rain, trouser legs dripping water."
Cause: The character card governs "what this person looks like by default"; being soaked and dripping are scene states that belong to each shot's visual description. Moreover, the reference images in the next lesson should follow the official recommendation of "even lighting, neutral expression"—wet hair and rain would skew the reference images.
Correct approach: Write only "short black hair" in the card; "soaked by rain, clinging to the forehead" goes into the prompts of specific shots.
❌ Starting generation with just the four words "delivery rider"
Symptom: The prompt contains only an identity word for the character; shot 02 generates a young man in his 20s, while shot 06 becomes an uncle in his 50s.
Cause: The identity word only constrains the broad category of occupational attire; age, build, and face shape are all left to the model's randomness.
Correct approach: The identity is only the first word of the fixed description block, followed by a full set of fields for appearance, clothing, and carried items.
❌ Writing a card for the daughter who never shows her face
Symptom: You carefully wrote the daughter's appearance and clothing, and even wanted her to show her face at the end.
Cause: Step 1's character list was skipped. The daughter appears only as an off-screen voice throughout the whole film (per the convention of the 2.2 script header), and the audience never sees her—writing a card is a waste, and having her show her face would also break the "few characters" scope control.
Correct approach: Circle the on-screen characters first before writing; mark the daughter in the character list as "off-screen voice only, never shows her face."
Advanced Tips (Optional)
Getting stuck on the personality field is normal—how do you draw "tough on the outside, soft on the inside"? You can hand this translation step to any conversational LLM. Expected result: the LLM translates each personality word into drawable features of "specific body part + degree," and you only do the acceptance check.
Below is the prompt designed for this course—use it as needed:
## Role
You are a film and television concept designer who excels at translating character personalities into drawable visual features.
## Task
Translate each of the character personality words I give you into appearance features that a generation model can draw directly.
## Input Information
- Character identity: [e.g., a delivery rider around 40 years old]
- Personality words: [e.g., dotes on his daughter, panics when things go wrong, tough on the outside but soft on the inside]
## Output Specification
A two-column table: personality word | drawable feature (expression, posture, habitual gestures, skin and build).
## Constraints
- ✅ Every feature can be drawn directly into a picture, specified down to the body part and degree
- ✅ The default expression stays neutral; do not translate into moments of emotional outburst
- ❌ Do not output abstract words like "kind" or "weathered"Before accepting everything wholesale, manually review it against the standard of Step 2: anything that cannot be drawn goes back for re-translation.
Lesson Summary
✅ This lesson delivers Asset 1: Character Sheet:
- Character list: only on-screen characters get cards—one card for Old Chen covers the whole film; the daughter is an off-screen voice only, no card
- Appearance, clothing, and carried items specified down to drawable details like colors and styles
- Personality is first written as words, then translated into expressions and postures in a neutral state—abstract words never enter prompts
- The fixed description block is within 100 characters and stored in a separate file—from now on, AI character settings reference the same block throughout the film
Character card field review:
| Field | How to fill | Old Chen example |
|---|---|---|
| Appearance | Age range, face shape and skin tone, hairstyle, build, distinctive features | Around 40 years old, square face, short black hair, short stubble |
| Clothing | Style + color, not a single item omitted | Yellow rider vest, dark blue long sleeves, black trousers, worn sneakers |
| Carried items | Items never out of hand throughout the film go into the fixed description block | White cake box tied with a red ribbon |
| Personality | After writing, translate into expressions, postures, and habitual gestures | Gentle eyes, corners of the mouth slightly downturned, takes quick small steps when anxious |
| Fixed description block | Within 100 characters, excluding weather/actions/shot numbers | 角色卡_老陈.txt |
Next Lesson Preview
The character card locks down what Old Chen is "written to look like"; the next lesson locks down what he is "drawn to look like": 3.2 Character Reference Images: Multi-View Generation Method—use the fixed description block to generate a set of multi-view reference images, and then every shot identifies him based on them.
You will learn:
- Why reference images lock in appearance better than plain text, and why they are still not an absolute guarantee
- The complete workflow for generating multi-view character reference images from the fixed description block
- Runway Gen-4 official reference image standards: even lighting, neutral expression, step-by-step iteration
- How to feed reference images together with the fixed description block to every shot