Chapter 77. Identity Drift Lab: From “It Looks a Bit Off” to a Fixable Variable#

Identity problems cannot rest on a reviewer saying the face changed in shot two. This lab generates Lin Xia under six high-risk conditions and uses the identity contract, the angle grid, state isolation and hard re-anchoring to locate where the drift comes from.

Figure 77-1 Identity stress testing must cover unflattering angles and performance conditions. State can be inherited; identity has to return periodically to the golden reference.

77.1 The stress matrix, separating state, and classifying drift#

The stress matrix.

The conditions are frontal still, left and right profile, looking down, strong backlight, open-mouth speech, and two-person occlusion. Each fixes wardrobe, hair, age, frame size and base reference, changing only one risk factor. Four candidates per condition, recording which anchors held and which non-identity variation occurred.

Separating identity, state, and style.

Inputs split into three layers. The identity contract governs facial proportion and stable features. Character state governs wet hair, emotion, injury and wardrobe. The style contract governs light, materials and lens character. Failing prompts commonly merge all three — describing a character as sad, rain-soaked, pale, exhausted and dishevelled at once — and a model may interpret exhaustion as a change of age.

Correct compilation locks identity first, then adds state: hold identity version lin_v03; add only 35 percent rain wetness at the hair ends and slight fatigue; age, face shape, mole and parting unchanged.

Drift is not one thing.

Morphological drift changes face shape and feature proportion. Feature drift loses the mole, the parting or the eye color. Age drift changes wrinkles and volume. Style drift makes the same person look like a different medium. Attribution drift swaps features between two characters. Cumulative drift comes from deriving each image from the previous generation until the error amplifies.

Each class routes differently. Style drift should not be fixed by raising facial similarity strength. Attribution drift needs slot and reference adjustments in multi-person shots. Cumulative drift needs a return to the golden identity rather than another patch on the latest frame.

77.2 Re-anchoring, feature migration, and the human threshold#

The re-anchoring experiment.

The same head-turn shot was built two ways. Chain A derived five times consecutively from the previous shot's end frame. Chain B returned to the golden three-quarter reference every second shot, carrying only the necessary state. By the fifth shot, A had moved the mole downward and widened the face. B stayed stable with a slightly stiffer motion connection.

The final strategy inherits state from the previous shot and re-injects identity from the golden reference; the end frame constrains the action relay without becoming the sole identity source. That is the two-track method: continuous state, re-anchored identity.

A feature migration incident.

Sharing frame with Lin Wei, a generation gave the mole below the left eye to Lin Wei. The root cause was not model stupidity. The reference collage placed the two portraits too close together, the prompt was organized by appearance paragraphs rather than frame slots, and the two costumes were similar in value in that version.

The fix declares foreground_left and background_right first, binds identity and wardrobe to each slot, states in feature ownership that the mole belongs only to Lin Xia, lowers the facial detail requirement for Lin Wei, and generates the characters separately for compositing where necessary. A fix cannot consist of adding an instruction not to mix faces.

Evaluation and the human threshold.

Automated similarity ranks and surfaces anomalies; a human approves based on anchors, character distinguishability and performance usability. Close-ups of principals carry a higher threshold than wides. Extreme expressions permit mouth changes and not skeletal drift. Low evaluator confidence routes to human review, and 0.89 versus 0.90 is not treated as a difference in kind.

identity_finding:
  shot_id: sh_e001_017
  condition: speech_closeup
  score: 0.88
  threshold: 0.93
  anchor_failures: [mole_missing, jaw_width_plus_7_percent]
  style_match: pass
  state_match: pass
  route: regenerate_from_gold_anchor

77.3 The repair ladder, the operating flow, and drift control charts#

The repair ladder and stop conditions.

Check for wrong references and mirroring first. Then reduce conflicting states. Then swap in the correct angle reference. Then lower motion and expression amplitude. Then add start and end keyframes. Then repair locally. Only then consider changing model or redesigning the shot. Each rung addresses its own class of problem.

Stop sampling when the same anchor fails three times running, when a repair renders the action unusable, or when coverage can substitute for the shot. Cutting to a rear view, hands, or the other person's reaction is not cheating, provided the narrative evidence stays complete.

SOP, checklist and exercise.

Establish the identity contract. Design the stress matrix. Separate identity, state and style. Classify by drift type. Compare cumulative derivation against hard re-anchoring. Run multi-person attribution tests. Combine automated and human evaluation. Work down the repair ladder. Add the incident to the regression set.

• Golden references cover the target angles, not only frontal.
• Temporary injury and fatigue are not written into the identity layer.
• Multi-person shots have slots and feature ownership.
• The end frame is not treated as the sole identity source.
• Every repair records the variable changed and the result.

Exercise: complete a six-condition stress test with four candidates each for one character, delivering identity_contract.yaml, stress_matrix.csv, anchor_findings.jsonl, drift_control_chart.png and repair_decisions.md.

How to use a drift control chart.

Do not plot average similarity alone. Put production order along the horizontal axis and record facial proportion, stable features, perceived age and human distinguishability separately on the vertical, marking model versions, reference packs and re-anchoring events. A stable average with the mole missing across three consecutive shots is still a systematic shift. A single very low shot may be a measurement anomaly caused by occlusion, so read it together with the visibility field.

The chart exists to reveal when the process leaves its stable region, not to grade a performer. If every profile drops on the same metric, improve profile references and the shot strategy. If a whole batch drops across all angles, check context compilation or a vendor change. Attributing errors to reproducible conditions is what upgrades a team from generating more candidates to controlling a process.

A note on sources#

This lab reconstructs a diagnostic method rather than reporting model behavior. What transfers is the stress matrix, the separation of identity from state, and routing each class of drift to the layer that can actually fix it.