Chapter 18. Character Visual Development#

18.1 Beautiful is not the same as reproducible#

The most common problem with a character concept image is that it only holds at one angle, one expression, one lighting setup. A real character asset has to remain recognizable in profile, in wides, in low light, while crying, while speaking, sharing frame with others, and in different costumes.

The goal of visual development is not to find the single most attractive image. It is to build a body of identity evidence strong enough to constrain everything generated afterwards.

18.2 Layering identity anchors#

Identity anchors divide into structural and decorative. Structural anchors are face shape, eye spacing, brow ridge, nose shape, jaw, hairline, build and height proportion. Decorative anchors are moles, scars, earrings, hair ornaments and color.

Decorative anchors are easy to check and cannot substitute for facial structure. A model that keeps the mole below the left eye while replacing the whole face has still failed.

identity_spec:
  id: CH_LINYUN
  age_read: 28-31
  face_structure: narrow oval, jaw tapering without a sharp point
  eyes: almond, slightly narrow, faint downturn at the outer corner, medium spacing
  brows: straight and fine, low arch
  nose: straight bridge, small but natural tip
  mouth: upper lip thinner than lower, corners level at rest
  skin: natural neutral tone, fine texture retained
  body: reads at 168cm proportion, level shoulder line, contained stance
  structural_forbidden: [round face, high brow arch, large upturned eyes, very pointed chin]
  symbolic: small mole below the left eye, the old pen

18.3 Characters must be distinguishable before anything else#

AI microdrama frequently ends up with every woman sharing one face shape, the same large eyes and the same pointed chin. Start development with a silhouette test: strip color and detail, keep only face, hair, shoulder line and height, and check whether the principals are still distinguishable.

Difference can come from age, facial geometry, hair volume, hairline, build, posture and the rhythm of movement — not only hair color. Lin Yun's silhouette is narrow and straight. Lin Wei's face is more heart-shaped, her hair waved, her shoulders and neck carried more openly. Zhou Lan's outline is stable, weighted by a chignon and darker masses.

18.4 From candidate exploration to identity lock#

Exploration may generate many faces and temperaments, and must not go straight into shot production. Screen candidates in three rounds: fit with the character logic; distinguishable from the other characters; stable under demanding angles.

Once selected, generate the standard pack: front, 45 degrees left and right, both profiles, rear of head, full body front/side/back, neutral expression, six core expressions, seated, walking, varied lighting, and the key costumes.

Approve only a small number of golden references. Everything else goes to a quarantine area and must never be retrieved automatically for generation.

18.5 What a turnaround is actually for#

A turnaround supplies costume, hair and body proportion. It does not guarantee facial identity. A white or simple background, consistent posture, complete full body and a wide enough frame all help extraction. Complex environments and artistic lighting bury the structure.

Front, side and back must belong to one design rather than three similar characters spliced together. Check ear position, hair length, shoulder width, costume seams and shoes by hand. Where necessary, stabilize each view separately before laying them out as a character sheet.

18.6 An expression sheet is not nine exaggerated faces#

Design expression sheets by intensity and by physical action unit: contained anger, faint suspicion, a brief flinch of hurt, a public cold smile, suppressed fear, genuine loss of control. Record eyes, mouth, brows, breath and head posture for each.

Test intensities one to three first — most performance in commercial microdrama happens in controlled close-ups. Reserve intensity five for a few climactic moments. Crying needs testing across welling tears, a single tear, broken breathing and full sobbing; a single contorted crying face is not enough.

18.7 Posture and motion signature#

Identity does not live only in the face. Lin Yun keeps her shoulders level, moves in small increments, and pauses before she acts. Lin Wei orients her body toward the audience in public and gestures more openly. Zhou Lan almost never turns her head quickly; she makes other people come to her.

These motion signatures go into the performance spec and the video prompts. Even when a wide shot cannot resolve a face, the audience recognizes people from silhouette and movement.

18.8 Lighting and angle stress tests#

A golden character must pass: frontal soft light, hard side light, top light, warm light, cool screen light, low illumination, backlit rim; and front, 45 degrees, full profile, slight high angle and slight low angle.

If a character is only stable in frontal soft light, do not approve them for a season. Fix the identity reference or adjust the route first. Keep at least one approved image per difficult condition for later reference.

18.9 Multi-person frame tests#

Being stable separately does not mean being stable together. Test the lead against the antagonist and against the romantic lead for height, facial independence, costume ownership and eye direction. Check whether the model has swapped moles, hair ornaments or colors between them.

Golden two-shots are only for space and proportion; they never replace each character's own golden single. During generation, always carry the single-person identity anchors back in, so that a two-person result cannot contaminate the assets in reverse.

18.10 Choosing golden references#

A golden reference must have the most accurate identity, a usable angle, neutral light, no hand or background errors, and clear provenance and commercial rights. More is not better; mutually contradictory references make a model average out a new face.

Record each image's purpose and its prohibited use: face_front_neutral governs identity; fullbody_grey_suit governs proportion and costume; banquet_cold_rim governs that specific light. Do not use an image carrying an extreme expression for every shot.

18.11 Identity similarity and human judgment#

Automated face embeddings catch obvious drift and cannot replace an art director. They may favor matching angles and expressions, and may ignore perceived age, bearing and local structure. Review structure, age, build, hair and symbolic objects together.

Establish grades: full pass; locally repairable; usable only in wides; rejected. Never add a barely-similar keyframe to the golden library.

18.12 SOP for character visual development#

First, extract age, status, living conditions and manner of movement from the character logic. Second, define structural and decorative anchors. Third, design silhouette differences against the other characters. Fourth, explore candidates and quarantine them. Fifth, complete the turnaround, full body, expression and posture packs. Sixth, run lighting, angle and two-person stress tests. Seventh, review with human judgment assisted by similarity scoring. Eighth, approve a small number of golden references and record their purposes. Ninth, write character ID, version and rights into the registry.

18.13 Fault tree#

Symptom: the mole is still there and the face has changed. Over-reliance on decorative anchors, with no structural description and no multi-angle golden images.

Symptom: every woman looks like the same person with different hair. Each character was pursued toward "attractive" separately with no silhouette differentiation first. Design the whole cast together.

Symptom: profiles collapse entirely. Identity was developed only frontally. Complete full profiles and make them an approval gate.

Symptom: crying scenes suddenly age or de-age the character. Expression stress testing was insufficient. Build graded expression sheets, and never use an extreme generation as an identity reference.

Symptom: features swap between characters in shared frames. Only two-shots were referenced, or a bad shared frame was added to the assets. Always combine with single golden images and quarantine failures.

18.14 Checklist, exercises and deliverables#

Check that structural anchors are specific; that principal silhouettes are distinguishable; that the turnaround is consistent; that expressions are graded; that posture and movement have signatures; that low light and profile pass; that shared frames do not swap features; that golden images are few and consistent; and that failures are quarantined.

Exercise one: design three distinguishable characters using black silhouettes only. Exercise two: write structural and decorative anchors for one character. Exercise three: build a six-lighting, five-angle stress matrix. Exercise four: from twelve candidates, select only six golden references and justify each purpose.

Deliverables for this chapter: the identity spec, the full-cast silhouette sheet, turnarounds, the expression sheet, pose and motion signatures, the lighting and angle matrix, the two-person proportion test, and the golden reference manifest.

18.15 Cast scoring must be done jointly#

Candidates cannot be chosen one at a time in a vacuum. Put the principals on a cast board first and check differentiation across seven dimensions: silhouette, perceived age, facial geometry, hair quality, costume color mass, height and motion signature. If two characters are each excellent yet easy to confuse in a vertical medium shot, at least one has to change.

Score in three categories. Narrative fit asks whether this face carries the character's class, history and sense of danger. Production fit asks whether it is stable across angles, expressions, lip sync and shared frames. Ensemble fit asks whether the whole cast forms a legible set of relationships. The final choice is not the highest individual score — it is the best combination once every blocking threshold is met.

Among the early candidates for Backlit Takeover, Lin Yun and Lin Wei were both designed with long hair, narrow faces and cool pale skin. Each read as urban professional in isolation and they looked like sisters in frame. The resolution kept Lin Yun's straighter bone structure and a low-centre-of-gravity hairstyle, and made Lin Wei rounder, more finished, lit with pearl highlights. The difference comes from structure and posture, not from one garment that might be changed.

18.16 Character assets must cover occlusion and distortion#

Real shots are not permanently unobstructed. Characters get partly hidden by hair, hands, documents, foreground shoulders, rain, motion blur and extreme light. Stress testing must include re-emerging from occlusion, because models frequently rebuild a different face in the occluded region. The goal is not that features stay legible under the obstruction — it is that identity recovers correctly once it clears.

Define the limits of perspective distortion as well. Slight high or low angles and close wide-angle framing change facial proportion, and must not change age, bone structure or recognizability. If the project's style uses close vertical close-ups often, build approved references for that perspective. If that perspective always breaks identity, write it into the forbidden camera positions rather than pushing the risk onto whoever generates each shot.

A hand against the face, wiping tears while crying, or leaning close to another person all challenge identity, hands and occlusion at once, so build static posture packs for them first. The video model is responsible for short movement, not for inventing a complex configuration for the first time.

18.17 Compiling a shot reference pack from the character library#

A complete character library may hold dozens of references; each shot should receive only a few, with clear responsibilities: one identity angle, one full-body image of the current costume, one expression or posture image, plus a special lighting test when needed. Feeding everything in at once makes the model average face shapes, blend costumes, and ignore what was actually authoritative.

The reference selector reads shot angle, size, emotional intensity, styling state and interaction type, and outputs a render context. Tag every reference with controls and does_not_control: a frontal identity image controls bone structure, not the current wetness of the hair; a wet-costume image controls material, not the face; a two-shot controls proportion, not individual identity.

Once a shot passes, its output must not be promoted automatically into the character references. It enters the golden library only through human approval, and only when the structure is accurate, there are no local errors, provenance is traceable, and it fills a clearly identified asset gap. Otherwise downstream references turn an incidental error into the new standard.

A note on sources#

Turnarounds, fixed identifiers and a principal asset library are frequent practices in Chinese AI comic-drama tutorials. This chapter emphasizes structural anchors, stress testing and quarantining the golden references, because a turnaround alone does not solve identity drift.