Help guide · 9 min read · reviewed 18 November 2026

Pavo AI Video: Keeping One Face Across a Shot List

Consistency is mostly bookkeeping. The model is only half the problem.

Short-drama workflows live or die on one question: is this the same person in shot three as in shot one? Most advice answers with a feature list — reference assets, character libraries, model choices. That helps, but the part creators actually lose time on is the record keeping. Without a record, every retry is a fresh guess.

The anchor sheet

Before generating a sequence, write down five things about the character in plain words. Not a paragraph: five short, repeatable phrases that you can paste unchanged into every prompt.

AnchorWrite it like thisAvoid
Faceoval face, straight hairlinebeautiful, striking, detailed face
Hairshort dark hair, parted leftstylish cinematic hair
Wardrobeplain cobalt-blue coatdeep rich navy layered outfit
Signature detailone visible collar notchinteresting accessories
Lightsingle window light from the leftdramatic studio lighting

Flat colour names matter. Rich, deep, dark and moody are interpreted differently on each render. One flat colour word used identically in every shot is far more stable.

Chain the outputs, not the inputs

The most reliable pattern is to use the end frame of shot one as the reference for shot two, and the end frame of shot two as the reference for shot three. You are not asking the model to remember an idea; you are giving it the exact pixels it just produced. Each shot then drifts at most a little, instead of diverging from a common origin.

  1. Approve one still. Save it as shot-01-approved.png.
  2. Generate shot one with a held end frame. Export that final frame as shot-01-end.png.
  3. Build shot two from shot-01-end.png, using the identical anchor sheet.
  4. Repeat. If a shot drifts, go back one step rather than back to the original still.
  5. Keep a one-line log: still used, prompt version, model, and the single detail that failed.

Design shots that hide the hard parts

You do not have to solve identity drift head-on. Choose coverage that avoids the failure: a wide shot with the character small, a close shot from behind, a reaction shot of another subject, a cutaway to hands or an object. Classic coverage is also a practical generation strategy.

Two-person frames are the hardest case. If both faces start to converge on the same appearance, split the scene into two single-subject clips and cut between them. The viewer reads a conversation from eyelines and timing, not from both faces sharing one generated frame.

Reviewing for drift

Put the end frame of each shot in a single row and look at them together at the same size. Check hairline, collar shape and ear visibility. Then play the sequence at normal speed. Drift that is invisible in stills often becomes obvious when the cut lands.

For a worked example, see the three-shot walk recipe, which uses one approved still, three locked setups and a held end frame on each shot.

Next

Turn the checklist into a render.

Open the studio with a still, a six-line brief and a held ending, then download the clip. Change one line per retry.