AI Video Consistency Explained: Why Characters, Scenes, and Props Drift

Maosika Editorial ｜ Last updated 2026-07-25

Characters change face, outfits swap between cuts, and props appear or disappear because every AI video clip starts from a fresh prompt with no memory of the last shot. The fix is an asset pipeline, not a better single prompt.

Why AI video drifts, in one paragraph

Every time you send a prompt to a video model, it is generating a new image of the world, not continuing the last clip. If the only thing connecting shot 2 to shot 1 is a text description like "a woman in a red dress," the model will freely reinterpret the face, the dress, the lighting, and the room. Drift is not a bug in the model. It is the absence of a continuity system between shots.

Consistency in AI short drama is the result of an asset pipeline, not a magic prompt. A production that treats every clip as an independent generation will get independent-looking results. A production that locks look-dev first, then binds every shot to those assets, gets a show that holds together across episodes.

The three types of drift, and what causes each

Drift type	What the viewer sees	Root cause
Character drift	Face changes, age shifts, outfit swaps, hairstyle resets	Text re-describes a character that already has a reference image; no locked character sheet
Scene drift	Rooms re-arrange, day/night flips, furniture moves	Scene described in words each time instead of referenced from a single location plate
Prop drift	A phone becomes a notebook, a jade pendant vanishes, a weapon changes shape	Props are mentioned in dialogue but never pulled into a visual catalog; shots bind the wrong reference or none

There is a fourth, quieter failure: cross-episode drift. The same character looks correct within episode 3 but unrecognizable by episode 12. This happens when there is no central record of who the character is, what they look like, and what state they are in right now (injured, disguised, revealed as heir).

The continuity bible, not model memory

A continuity bible is a structured record of the entire production: character identities, stable traits, current state (wounds, disguises, status changes), relationships, open and resolved plot threads, an appearance schedule by episode, batch-level plot notes, and visual descriptions of recurring props. In traditional writers' rooms this is a physical or shared document; in an AI pipeline it must be a structured asset the system reads from before writing or shooting each batch.

The key principle: do not rely on model memory; rely on a structured archive. A language model reading a 200-episode script will forget details. A system that pulls only the relevant slice — current character state, unresolved threads, recent batch summary, this batch's main arc — will not.

Look development before shooting

The single most common mistake teams make is going from script straight to video generation. In live action you would never roll camera before costumes, makeup, and set are locked; AI production is no different.

A proper look-dev stage produces three things before a single clip is generated:

Character sheets for every named role, in the chosen visual style, with multiple views and a locked face anchor.
Scene plates for every recurring location, empty of people, showing the room, lighting, and layout.
Prop catalog items extracted from the script using the script's own wording — not re-imagined labels.

Style must be locked once and reused everywhere. If character art is generated in a 90s anime style but video prompts drift into realistic film language, the result is a show that changes genre between the poster and the first frame. A production-ready system keeps character art, scene art, prop art, and video style tags on the same style path.

The reference image map per shot

Once look-dev is locked, every shot needs its own reference table, assembled in a fixed order: scene plate → props in this shot → character sheets. Slots are only filled when an image exists; empty slots are marked as text-only rather than silently fabricated.

The highest-priority rule, worth stating on its own:

When a character has a reference image, the prompt must not re-describe their clothing or appearance in text. Text only describes action, expression, and injury state.

This rule directly targets the most common AI failure mode: text and reference fighting each other. If the reference shows a woman in a black suit and the prompt adds "wearing a red dress," the model has to negotiate two incompatible instructions, and the output is a coin flip. Remove the text description; let the reference own appearance.

Other reference-map disciplines that prevent drift:

Manual character binding. A script that calls someone "the officer" must be bound to the correct character sheet for "Li Qiang," not left to the model to guess.
Prop include/exclude control. A scene set in a cafe should not drag in a sword from episode 5 just because it exists in the catalog.
Era routing. For time-travel or flashback stories, pick the correct costume version of a character based on scene location keywords, not the default modern look.
Empty scene plates. Scene reference images must contain no people, so a character sheet can be composited in without a ghost extra in the background.

Shot blocks, not whole episodes

A vertical short drama episode is typically 1–2 minutes. Trying to generate it as one long clip guarantees drift, because the model has to hold character, scene, and plot across too long a span with no checkpoints.

The industrial approach is to cut each episode into shot blocks of about 10 seconds each, with a soft cap around 200 characters of script per block. Long actions are split on beats, pauses, or sentence boundaries. The creator sees not "Episode 3, one video" but "Episode 3 → Shot 1, Shot 2, Shot 3…" — each with its own prompt, its own reference map, its own output file, and its own history of takes.

This is how live-action editing rooms work: a clip list, not a single master take. It is also how consistency is maintained at scale, because each block is small enough to be re-shot without redoing the whole episode.

Engineering shot prompts like a storyboard

A video prompt is not a paragraph of prose. It is a shot list instruction. A production-grade prompt has eight components:

Precise subject — who is in frame, named and bound to reference.
Action detail — specific, quantified, low-velocity motion; avoid explosive motion that breaks the model.
Scene environment — location, time of day, weather, drawn from the scene plate.
Lighting and color — motivated light sources, tonal palette.
Camera movement — one move per shot; no combined push-pan-tilt in a single clip.
Visual style — tags inherited from the locked style path, not re-decided per shot.
Image quality — resolution, stability, facial consistency anchors.
Constraints — no watermarks, no logos, no twins in multi-character shots, style anchors for non-realistic work.

Complex scenes use a three-part structure: overall setup, then shot 1 / shot 2 / shot 3 as numbered beats, then the constraint package. Shots are numbered, not timestamped — writing "0–3s" in a prompt is fragile; "Shot 1:" is not.

Two formatting conventions keep audio from drifting:

Dialogue is wrapped in {curly braces}.
Sound effects are wrapped in <angle brackets>.
BGM notes are wrapped in (parentheses).

The prompt for a shot only sees that shot's assets. It does not get the whole script, and it does not get props from other scenes. Cross-shot contamination is a major source of drift, and it is solved by feeding the model only what belongs in this block.

Multi-take selection is human work

There is no reliable automatic scoring for AI video takes. A system that claims to auto-pick the best clip is guessing. The honest production workflow is:

Generate a take from the locked prompt and reference map.
Review it against character, scene, prop, and performance expectations.
Either accept it, edit the prompt and re-shoot, swap a reference image and re-shoot, or cut the block differently.
Keep the history of takes per block so you can go back to an earlier version.

This mirrors a real set: the director watches a take, gives notes, changes the setup, and rolls again. The tool's job is to make re-shoots cheap and traceable, not to replace the director's eye.

Where the approach does not work

It is worth stating the boundaries plainly, because teams that expect miracles waste the most time:

Reference images are not a hard gate. You can skip them and go text-only, but quality is usually worse; professional workflows lock look-dev first.
Shot blocks are a heuristic, not timecode-accurate editing. A long action may still produce a slightly long block; fine cutting and stitching are a post-production step.
There is no cross-shot auto-extend workflow. The unit of production is a single shot block with its own references and output; automatic continuation across clips is not a reliable feature.
Character consistency depends on the quality of the character sheet. The pipeline routes the right image, but a poorly generated sheet will still produce inconsistent results.
Reference tokens in prompts are guided by convention, not hard-stitched. Creators should still read the prompt before shooting and confirm references are present.

These are not failures to hide. They are the shape of the problem. A tool that pretends they do not exist will let a team down in episode 10, when drift has already compounded beyond repair.

A production checklist

Before you generate your first clip of an episode, confirm:

[ ] The continuity bible is updated with current character state and open threads.
[ ] Every named character in this episode has an approved character sheet in the locked style.
[ ] Every recurring location has an empty scene plate.
[ ] Every prop mentioned in this episode has a catalog image, or is explicitly marked text-only.
[ ] The episode is cut into ~10-second shot blocks, not treated as one long generation.
[ ] Each block's reference map is assembled in order: scene → props → characters.
[ ] Prompts do not re-describe appearance for characters that have reference images.
[ ] Camera movement is one move per shot, numbered by shot, not timestamped.
[ ] You have budgeted for multiple takes per block and time to review them.

High-quality vertical drama is never one long generation cut up in post. It is a stack of controlled units, each small enough to correct, each bound to shared assets, each reviewed by a human before it ships. Consistency comes from the pipeline, not from luck.

About Maosika

Maosika (猫斯卡) is an AI production operating system for vertical short dramas. It encodes the workflow described above — intake scoring, a locked creative brief, a structured story archive, look-dev for characters/scenes/props, ~10-second shot blocks, engineered shot prompts with eight components, a reference-map system that enforces "reference owns appearance," and a multi-take shooting queue — as a single pipeline from idea to finished footage, with human sign-off at every stage that matters. It does not claim one-click hits or zero-drift output; it turns the repetitive, drift-prone parts of production into a constrained assembly line, while aesthetic judgment stays with the creator.

About Maosika — Maosika · Professional AI Video Production System. It connects briefing, scripting, look development, shot prompts and delivery into one reviewable pipeline. www.maosika.com