Track consistent characters in AI video across every shot

Create a character record before comparing shots

Give the two partners stable names such as A and B. Attach an approved reference for each and write a few visible identifiers: hair silhouette, face proportions and a recognizable accessory when one exists. Describe what the image actually shows. Do not infer hidden features or add personality labels that cannot help a visual comparison. This page begins after photo selection and role assignment. The photo guide owns the starting portraits; the multiple-face workflow owns mapping identities inside a replacement operation. Here the question is whether the same people remain recognizable across an edited sequence, including shots produced in separate attempts.

Vendor reference tools can help guide character appearance, but documentation of a reference feature does not prove perfect continuity in an output. Dreamina's character-consistency page describes a reference-oriented workflow. Treat the continuity sheet as an independent review record that remains useful whichever external tool produced the clip.

Separate stable identity from intended transformation

For a zombie story, create distinct present and memory state records. The present may contain altered skin and worn clothing; memory may show healthier appearance and earlier wardrobe. Record those intended differences beside the stable identity features. Otherwise a reviewer may reject the transformation itself or excuse every unexpected face change as part of it.

Be precise about the boundary. A skin-color change may be planned; a different nose outline or hairline may not be. If an eye is hidden by shadow in one shot, mark it unobservable rather than passing it by assumption. The sheet should distinguish visible evidence from information you cannot judge.

Keep A and B separate even when they share a story state. “Both zombies” does not define which jacket or accessory belongs to whom. Swapped wardrobe can weaken recognition before a face swap becomes obvious. Each character needs an individual row with the state-specific identifiers that should survive the next cut.

Build a shot grid with comparable evidence

Create one row per shot and columns for shot ID, state, A's appearance, B's appearance, wardrobe, screen direction, usable range and decision. Add a representative still, but do not let it stand in for the entire moving clip. The still makes the grid readable; the range note tells you which action was inspected.

Choose comparable views where possible: a clear face before a turn, the last clear face before overlap and the first clear view afterward. A frontal reference and an extreme profile will naturally show different shapes. Compare the features visible in both rather than forcing an exact outline match across incompatible viewpoints.

Label every still with its source clip and position. “P3 after overlap” is easier to find than a loose screenshot called image-final. If you later shorten the clip, update the usable range in the sheet. An accepted still from a discarded part should not become evidence for the revised shot.

Inspect identity through crossings and reappearance

Watch each shot normally, then pause around occlusions: a face turns away, one partner crosses in front of the other, or a hand covers a cheek. Compare the last visible evidence before concealment with the first visible evidence afterward. A character can look correct at both ends of a clip yet change identity during the middle.

Trace clothing ownership during a crossing. The screen-left person may become screen-right while still wearing the same coat. Update screen position in the direction column instead of renaming the character. If both faces become ambiguous, mark the interval as uncertain and inspect whether the edit can avoid it without breaking the action.

Check small background appearances too. A secondary reflection or distant partner can introduce a duplicate or inconsistent face. Whether it matters depends on visibility in the final crop, but it should be a conscious decision. Enlargement or sharpening may make a previously unobtrusive inconsistency more apparent later.

Compare neighboring shots as a continuous event

After the individual shot review, inspect every join. A looks toward screen-right in one shot; the next view should make that relationship understandable. A sudden direction reversal may be a deliberate camera change, but it needs context. Record the camera or staging reason rather than treating direction as an absolute identity rule.

Compare wardrobe at the story-state boundary. If the memory intentionally changes clothes, make the time shift clear enough to explain it. If two adjacent present-day shots change a jacket without explanation, choose a compatible replacement or simplify the sequence. A color grade cannot reliably restore a missing garment shape.

Review the embrace transition as a special join. Identity, body overlap and transformation may all change near the cut. Use the match-cut guide for the movement anchor, and use this sheet to confirm that the people carrying that anchor remain the intended pair. A geometrically smooth cut can still connect two inconsistent characters.

Choose a repair that preserves the visible event

Classify a failure before choosing a repair. A short unusable tail may be removed if the remaining action is complete. A wrong identity through the main gesture usually requires a replacement shot or a reworked external generation. A wardrobe mismatch across a memory boundary may be resolved by making the temporal change clearer if the difference was intended.

Do not use tighter cropping as an automatic repair. It may hide the wrong jacket while removing the second partner or the hand that explains the embrace. Evaluate the repaired shot for the same narrative purpose that justified including it. A technically clean close-up can fail to show the interaction the story needs.

Return to the baseline before another attempt. Changing references, wardrobe language and camera direction together makes the next result hard to diagnose. Preserve the accepted identity record and revise the specific input or instruction related to the observed defect. Label any newly proposed state so the sheet does not quietly change its acceptance criteria.

Review the accepted sequence at final viewing size

Once repairs are complete, watch the actual edit from beginning to end with the continuity sheet nearby. Confirm that accepted ranges are the ranges on the timeline. A later trim can reintroduce a rejected reappearance or remove the cue that made a wardrobe change understandable.

Inspect the final vertical or wide export at ordinary viewing size and in closer view where a concern exists. Record unresolved ambiguity explicitly. “B unobservable behind A for this interval” is a stronger note than a blanket claim of perfect consistency. Decide whether that ambiguity is natural concealment or prevents viewers from following the relationship.

Keep the sheet with the project and update it when replacing a clip. The Zombie Couple site's current local preview does not perform this across-shot review or generate a finished film. The practical endpoint is your own accepted sequence with traceable evidence for both characters, including deliberate transformations and any limitations you chose to retain.

• Every accepted range points to the actual timeline clip.
• Planned state changes are recorded separately from defects.
• Both identities are checked after concealment.
• Wardrobe ownership survives crossings.
• Final framing leaves enough evidence to recognize the pair.