Chapter 22. Character Identity Consistency#

22.1 Identity consistency is a chain, not a setting#

Fixing a seed, repeating a character description, or referencing the previous frame each removes some randomness. Reliable identity comes from five layers constraining the result together: approved identity assets, a reference suited to the current angle, stable styling state, a reviewed still keyframe, and a second review after motion generation.

Any one layer failing can change the face. A continuity system has to record which layer failed. It cannot simply log "generate it again."

22.2 Four layers of character information must stay separate#

The identity layer is facial structure, age, build and permanent markers — things story does not casually change. The styling layer is hair, costume, makeup and accessories. The story state layer is wounds, wet hair, fatigue, stains and emotion. The shot layer is angle, size, light, action and composition.

Write all four into one long prompt and changing a costume easily changes the identity by accident. The system should reference them separately:

Figure 22-1 The character identity reference board. The top row fixes frontal, forty-five degree and profile structure under consistent light, lens and head scale. The bottom row changes only story states such as wet hair, fatigue and injury, without redefining the person. The detail images on the right are used to review identity anchors: brow shape, mole, jaw and hairline.

The point of a reference board is not collecting more attractive photographs. It is assigning each image an explicit authority. The top row may determine bone structure and facial proportion; the bottom row may only show how one state acts on that same identity. Even when a state image is more attractive than an identity image, it must not be promoted to a new golden identity — otherwise the rain, fatigue and injury versions start contaminating ordinary shots in reverse.

character_render_context:
  identity: CH_LINYUN_FACE@v03
  body: CH_LINYUN_BODY@v02
  wardrobe: WD_LINYUN_GREY_SUIT@v02
  state: CHSTATE_E001_B04
  shot:
    angle: right_three_quarter
    size: medium_close_up
    lighting: LOC_BOARDROOM_DAY_key

22.3 Choose references by task#

A frontal face suits frontal identity. A profile must reference an approved profile. Full-body shots need body proportion and costume references. Extreme emotion needs an expression reference at a similar intensity. Do not push every golden image indiscriminately into each generation.

Too many mutually contradictory references make a model average out features, age and costume. Build a reference_selection per shot stating what each image is responsible for.

22.4 What a seed actually does#

A seed helps reproduce a random starting point under the same model, parameters and a similar composition. Once angle, pose, prompt, model or resolution changes, it cannot guarantee identity. Treating a seed as an identity ID fails the moment the camera moves.

Seeds belong in the generation ledger for reproducibility. Identity review still judges against golden references and structural features.

22.5 Anchor on stills first#

Generate a still keyframe for every shot first. Review facial structure, age, hair, build, costume and markers. Only an approved still proceeds to video.

Video models can drift during motion, so a passing first frame does not mean the clip passes. Check first, middle and last frames, plus head turns, occlusions and expression peaks. If drift appears only at the end, shorten the shot or cut earlier. If it is wrong from the first frame, return to the keyframe.

22.6 Accumulated drift and re-anchoring#

Using the last frame of the previous shot as the next reference helps action continuity, and it also passes small errors forward. After ten links in a chain, a character can gradually become a different face.

The correct method is dual reference: the previous end frame governs action and composition, while the original golden identity image governs the person. Every few shots, or whenever the angle changes, regenerate from the golden asset so the chain does not recurse indefinitely.

reference_policy:
  continuity_ref: E001_S08_endframe
  identity_ref: CH_LINYUN_FACE_LEFT45@v03
  wardrobe_ref: WD_LINYUN_GREY_SUIT_FRONT@v02
  continuity_weight: medium
  identity_priority: highest
  hard_reset: false

22.7 Profiles, wides and occlusion#

Profiles vary a great deal, so check the nose-to-forehead line, jaw, ears and hairline. Faces in wides have too few pixels to demand accurate features; maintain identity through build, hair, costume silhouette and motion signature. When someone re-emerges from occlusion, models tend to rebuild a new face, so re-inject identity at the keyframe once the obstruction clears.

A wide can be marked "identity indeterminate, silhouette passes." Do not force a low-resolution face into a waxwork.

22.8 Extreme expressions and lip sync#

Sobbing, shouting, closed eyes and a wide-open mouth all change facial geometry. Identity review needs expression tolerance: it cannot demand pixel equivalence with the neutral face, and it cannot accept a wholesale change of age or face shape.

Generate stills using approved expression references first, then add short motion. Split long dialogue into short lip-sync shots with off-screen continuation to reduce sustained deformation.

22.9 Repair paths#

A slightly wrong mole, skin tone or stray hair can be repaired locally. Wrong facial structure means redoing the keyframe. Brief drift during motion can be shortened or covered with an insert. Drift across the whole clip means regenerating the video. Errors across many shots mean returning to the identity asset or the model route.

Face swapping is not the default answer. It can correct a frontal view while creating new problems in profile, lighting, skin quality and edges — and it carries likeness and rights risk.

22.10 Identity scoring#

Check facial structure, perceived age, hair, build, permanent markers and overall recognition separately. Hero close-ups have a high threshold; background wides permit a different tolerance.

identity_qc:
  shot: E001_S09
  face_structure: 5
  age_read: 5
  hair: 4
  body: not_visible
  symbolic_marker: 5
  overall: pass
  reference: CH_LINYUN_FACE_FRONT@v03
  reviewer_note: slightly more loose hair on the right; identity unaffected

Automated similarity is for ranking anomalies and cannot approve on its own. A cross-angle score may be low while the art director can still confirm it is the same person.

22.11 SOP for identity consistency#

First, lock a small set of golden identity and angle references. Second, separate identity, styling, state and shot layers per shot. Third, select reference responsibilities by angle. Fourth, generate and review the still. Fifth, generate short video from the keyframe. Sixth, check first, middle, last and expression peaks. Seventh, dual-reference the previous frame and the golden image. Eighth, hard re-anchor periodically. Ninth, route errors by layer to local repair, new keyframe, new video, or new asset.

22.12 Fault tree#

Symptom: frontal is right, profile is wrong. There is no approved profile structure and the frontal reference is being forced onto a profile. Add golden profiles.

Symptom: it looks less like her the further you go. Only the previous frame is being referenced and drift accumulates. Return to the golden identity and shorten the recursion chain.

Symptom: the still is right and the face changes in the video. Motion, occlusion or lip sync are causing dynamic drift. Shorten the motion, reduce the expression amplitude, or regenerate.

Symptom: the face matches but the age keeps moving. Identity review is ignoring age and skin. Make perceived age its own blocking item.

Symptom: the repair looks pasted on. A local face swap does not match lighting or skin quality. Redo the keyframe, or unify the composite edges and grade.

22.13 Checklist, exercises and deliverables#

Check that the four layers are separated; that references match the angle; that seeds are used only for reproduction; that stills pass first; that first, middle and last frames of video are reviewed; that golden and continuity references are both carried; that wides are judged on silhouette; and that repairs return to the correct layer.

Exercise one: design frontal, profile, wide and crying reference packs for one character. Exercise two: simulate five shots of recursive drift and set the re-anchor point. Exercise three: route ten identity errors to assets, keyframes, video or compositing.

Deliverables for this chapter: the character render context, the reference selection, the re-anchoring policy, still and motion identity QC, the drift log, and the identity repair decision tree.

22.14 The identity contract: define first what may never change#

"Looks like the same person" is still too vague. Each principal needs an identity contract splitting identity into blocking items, strong preferences and permitted variation. Blocking items usually include facial bone structure, age range, eye spacing, the nose-to-forehead relationship, jaw outline and permanent markers. Strong preferences include skin tone range, brow shape, hairline and build. Permitted variation includes muscle changes from expression, temporary makeup, wet hair and reasonable color shifts from lighting. A reviewer decides which layer a deviation belongs to before deciding whether to reject it.

identity_contract:
  character: CH_LINYUN
  immutable:
    craniofacial: long oval face, relatively high zygomatic arch, narrow jaw
    eyes: inner double eyelid, spacing about one eye width
    nose: straight bridge, tip drawing slightly down
    age_read: 28_to_32
    marker: small mole below the left eye
  strong_preferences:
    skin_range: neutral_light_03_to_05
    build: slim_athletic
    hairline: center_soft_peak
  allowed_variation:
    - natural change in expression musculature
    - slight swelling on the right side after injury
    - skin tone shifts from different color temperatures
  forbidden:
    - juvenile round face
    - the mole moving to the right side
    - an age jump with no story event

An identity contract cannot merely sit in the character bible for people to read. The keyframe agent compiles blocking items into positive and negative constraints; the review interface puts blocking items first; automated similarity only flags possible anomalies. When a character genuinely must change — surgery, ageing, disguise — create a new identity state version and write the story event linking old to new, rather than overwriting the original asset.

22.15 The angle grid and the expression envelope#

Three images — front, left, right — are nowhere near enough. Principals need an angle grid: frontal, 30–45 degrees each side, near profile each side, and any necessary high and low angles. Then build an expression envelope for the frequent expressions: contained smile, suppressed anger, checked surprise, pre-tears, and vowels during speech. The angle grid answers where the structure is seen from; the expression envelope answers how far the structure may deform. With both, the model does not have to guess every extreme state from one neutral frontal image.

You do not need a golden image at every intersection. Derive the requirement from the shot list: if the season has almost no full right profiles, do not invest equally there. If the lead has a dozen scenes of suppressed crying, then pre-tears, the tear line appearing, and loss of control are three states worth approving in advance. Asset priority comes from shot coverage and the cost of failure — not from completing a character turntable that merely looks thorough.

The angle grid also stores shooting conditions: focal length tendency, camera height, light direction and whether it suits lip sync. A profile generated with a low wide-angle camera cannot serve as universal identity truth, because perspective has already altered the geometry. A golden image only becomes a production asset once its scope of application is declared.

22.16 A drift control chart for batch production#

One shot passing does not mean the episode is stable. Build a contact sheet per character in story order, each cell using a similar crop and labelled with shot, angle, state version and identity score. Reviewers scan the whole sheet for two seconds first, looking for the cell that suddenly appears to be a different performer, then go into detailed comparison. Fast overall recognition is closer to how the audience perceives identity than zooming into each image in turn.

Figure 22-2 Continuity frames supply action and composition; the angle grid and expression envelope supply the permitted range of deformation. Once a sequence keeps departing from the golden identity, a hard re-anchor is required — you cannot keep recursing on an end frame that has already drifted.

The diagram carries two opposite risks. Returning only to the golden identity every shot breaks action and lighting. Attaching only to the previous frame lets identity drift slowly. A production system must carry both sources and weight them by shot function. A hard re-anchor does not mean returning to an identical frontal expression; it means returning to the same structure and then applying the current angle, expression and story state.

Record the drift trend, not merely pass or fail. If jaw width increases across six consecutive shots, the recursive reference is departing even while each shot marginally passes. Build control charts for structural sub-scores: three consecutive shots deviating in the same direction triggers a hard re-anchor; repeated failure at one angle escalates to a reference asset gap; two model routes producing two stable but different "performers" means choosing one, because they cannot alternate inside an episode.

Batch boundaries are natural re-anchor points. Every new scene, new look and new generation batch recompiles from the golden identity and approved state rather than inheriting the last image of the previous batch. The previous shot supplies only action and lighting continuity; identity authority always comes from the character asset library.

22.17 An identity incident from Backlit Takeover#

In the pilot, Lin Yun was stable in the frontal boardroom close-up, and turning to the left profile beside the screen shortened her nose and widened her jaw — she read as a younger, different performer. The team first tried repairing it with a frontal face swap, and the edges looked like a mask under the blue screen light while the mouth lost volume. The root cause was not insufficient swap strength. The character library simply had no approved left profile and no screen-light condition.

The correct fix ran in three steps: build a left profile structure candidate from the frontal and three-quarter golden images; approve the profile under neutral light and then produce a blue screen-light test; regenerate the still keyframe and restore short lip sync only after it passed. Every subsequent screen-side shot referenced that same profile version. The team also added "a new angle appearing for the first time under special lighting" to the asset gate, so the same class of problem could not spread across the season.

A note on sources#

Turnarounds, fixed features and reference images are the practical foundation, but a stable character needs reference selection, a still gate and a motion re-review. This chapter separates in particular the different responsibilities of the seed, the previous frame, and the golden identity.