AI Short Drama Character Consistency (2026): Why Faces Change Between Shots, and the Three Layers That Lock Them

What "consistent" means: three layers

1. Identity features: face shape, brows and eyes, nose and lips, skin tone, hair, build, apparent age, plus one instantly recognisable anchor (a scar, a mole, a piece of jewellery, a haircut). Carried by the character sheet.
2. Wardrobe, makeup and props: what this character wears and holds in this scene, and what the set contains. Carried by the per-character costume description from script analysis and by the scene plate. Costume is per character and per social class, not one style tag for the whole series.
3. Light and art direction: colour discipline, materials, set-dressing density, where the light source is. Carried by the scene plate and by the lighting lines in each shot. The reference images themselves must stay neutral.

Why faces drift: three root causes

Layer 1: the character sheet

The sheet exists for recognisability, not mood. Four hard rules:

• Pure white background, as its own clause. State it as a standalone constraint and repeat it in the negative list. Buried in another paragraph, it gets overridden.
• No flat light. "Even lighting, no shadows" is the recipe for a plastic face: flat light erases skin texture and bone structure. Write one directional key light at roughly 75 degrees with fill and rim, and ask for shadow under the nose and cheekbone, a catchlight, and gradual falloff.
• Identity first. Fixed prompt order: subject identity features → identity lock (no influencer face, no smoothing of distinctive features) → technical block → framing constraints. Script analysis produces very detailed appearance text (bone structure, moles, scars); it has to lead, not sit in the middle.
• Face > posture > costume, and the expression should act. Declare that this is an actor's look test, not an e-commerce outfit shot; forbid pulling back to show the outfit at the cost of the head. Do not write "neutral expression": that deletes the very quality (a blade-like stare, say) that makes the character recognisable.

Known failure: characters in long coats or robes get swallowed by the garment in a 9:16 full-body frame, the coat filling seventy percent of the image and the face shrinking. That is a physical limit of the framing. The fix is a separate half-body sheet that carries the face, at the cost of one more image per character.

In SceneMixer the appearance field produced by script analysis becomes the subject of the sheet prompt directly. Each character gets one 9:16 full-body sheet on white, with face and half-body crops cut automatically, and every later video generation attaches that character's references. Before judging sheet quality, confirm which image model produced it: output size and prior differ a lot between models, and the evidence is the image dimensions and the task record, not a guess.

Layer 2: the scene plate

A scene image is an empty "plate" the video model references. The standard is again neutrality:

• No people. Express scale with the building's own parts (a door two people tall, a table one arm wide).
• Three planes of depth. Foreground, mid-ground and background each with explicit materials and dressing, so the model has room to place a camera.
• Describe where the light comes from, not where it lands. Write "a warm-white pendant throws soft light down from upper right", not "light brightens the tabletop". Do not write a shadow terminator line, and avoid saturated colour words: 3200 K on skin is warm white, and "yellow" gets you yellow paint.
• No atmosphere grading. No mood grading, silhouettes, fog or light beams, and none of the old trigger words (cinematic, dramatic, 8K). Fog in the plate appears in every shot of that scene.

Art direction is a project-level setting applied once, at script analysis: it shapes every scene description and storyboard, and is never appended to the character sheet or any other image prompt, otherwise the white-background sheet gets lit like the set.

Layer 3: continuity in the storyboard

• Long segments, short shots. One segment is one generation call and should run close to the 15-second ceiling; raise pace by shortening individual shots, not by slicing the segment. Every extra segment restarts positions from a new first frame and throws away continuity.
• Carry over inside the segment. Where the character sits, what they hold and what they were doing at the end of the previous shot is where the next shot starts.
• Only write who is in frame. Never name absent people. "Mark exits" or "Anna is not present" raises the chance they appear, because nouns inside a negation act as attractors.
• Lighting by cause. Same rule as the plate. Light position, medium and grade each stated once, with no repeats and no conflicts.

Reference inputs by model

Only verified cells are filled; blanks are blanks, not guesses. What decides consistency is not whether the cap is 9 references or 50, but whether the few you attach are neutral.

Acceptance checklist: five shots of the same character side by side

• ☐ Face shape, brows, eyes, nose and lips read as one person across all five.
• ☐ The visual anchor (scar, mole, jewellery, haircut) is present and in the same place in every shot.
• ☐ Within one scene, light direction and colour temperature match; no warm shot next to a cold one.
• ☐ Costume and hand props are continuous inside the segment, nothing appears or vanishes.
• ☐ No absent character walks in; no second "lead" appears.
• ☐ Skin has texture and gradual shading, not a flat plastic surface.

When a shot fails, go back to the layer it maps to and fix that layer. "Strengthen the prompt" is not a fix.

Further reading

How the script itself sets up consistency (one name per location, no more than three or four sets per episode) is in Vertical Short Drama Script Format and Hook Structure. Disclosure and likeness rules are in AI Short Drama Compliance Checklist 2026. To watch the method on a finished series, open the sample dramas: every one shows the cast sheet and each segment's script and video.

FAQ