Scene Memory Is Not Mood

Question

In human directing, a scene description can be loose.

A living room can be "warm." A meeting room can be "formal." An auction hall can be "luxurious." These words work because a human production team can translate mood into physical choices. A production designer can choose the floor. A cinematographer can choose the light direction. A director can remember where the sofa sits across cuts.

In AI video generation, mood is not memory.

A model can preserve the mood while changing the room.

That distinction became one of the most important lessons from repeated short-drama production. The render may still look cinematic, but the story world quietly mutates: the sofa changes, the tea table moves, the floor becomes glossy, a daytime room becomes night, audience chairs appear behind the wrong subject, or stage furniture leaks into an audience close-up.

The mistake is treating scene details as atmosphere.

They are not atmosphere. They are continuity tokens.

Production Basis

The pattern was consistent across different kinds of spaces:

• a family living room needed the same sofa, tea table, floor, broken cup, and phone position across multiple Copy Blocks
• a meeting room needed stable natural light, glass partition, long wood table, black leather chairs, and non-reflective floor behavior
• an auction hall needed separate host/stage, audience, aisle, and payment-counter zones
• red audience chairs were valid in one zone but unsafe when they appeared behind the host
• a display cabinet was valid near the stage but wrong behind a seated audience speaker
• exterior shots needed user-removed props to disappear completely, not only from the visible picture prompt

These were not failures of style. They were failures of scene memory.

The model had enough visual skill to create plausible rooms. It did not have enough external state to know which version of the room had to remain true.

Continuity Tokens

For repeated AI video scenes, these details should be treated as hard memory:

time of day
main light source
color temperature
light direction
floor material
wall or partition type
furniture color
furniture material
furniture shape
furniture position
active zone
allowed background anchors
forbidden background anchors

If these are left as mood, the model will reinterpret them. If they are compiled as scene memory, the prompt has a chance to preserve them.

This does not mean every public prompt should become a database dump. It means the scene compiler should decide which local facts need to be repeated for the current shot.

A repeated scene needs a memory packet:

Scene: living-room sofa zone
Time: daytime
Light: cool natural window-side light
Floor: matte, non-reflective
Furniture: dark wood tea table fixed lower center
Background: sofa rear layer, cabinet rear-right
Visible people: only if the current shot function requires them
Forbidden drift: no night, no extra family silhouettes, no new table form

That packet is not a beautiful description. It is a production contract.

A Location Is Not One Unit

The most dangerous word in scene prompting is often the location name.

"Auction hall" sounds specific, but it is too large. It may contain a stage, a host area, rows of audience seats, an aisle, a display case, a payment counter, and entrance lighting. All of those anchors may be valid somewhere. They are not valid everywhere.

If the compiler imports the whole location into every shot, Seedance can combine incompatible areas:

• host close-up with audience-seat background
• audience reaction with stage display cabinet behind the face
• payment beat with auction-stage furniture
• standing meeting-room proposal with chairs becoming active seating

This is why scene memory has to be zone-based.

The compiler should ask:

Which zone is active?
Which anchors are allowed in this zone?
Which anchors belong to another zone?
Which anchors should be suppressed from this local request?

The answer should be positive and local. A prompt should not rely on a long negative list. It should first define the current physical truth.

People Are Expensive Anchors

The same rule applies to visible people.

Background humans are tempting because they make a space feel populated and continuous. But people are expensive. They can invent reactions, pull focus, create mouth movement, change emotional geometry, or imply a relationship that the shot does not need.

In one production pass, extra family silhouettes were originally used as continuity aids. The intention was understandable: keep the room from feeling empty. But the effect was unstable. Multiple named bodies increased the prompt load and made speaker shots more fragile.

The stronger rule is:

Every visible person must have a shot function.

Allowed functions include speaker, closed-mouth listener, action partner, prop holder, spatial opponent, public crowd texture, or remote authority insert. Invalid functions include "keeps continuity," "shows they exist," or "fills the background."

Physical anchors should carry ordinary continuity first.

A sofa can anchor position. A table can anchor geography. A floor can anchor posture. A light line can anchor direction. A person should enter the frame only when the shot needs that person.

Crowd Texture Is A Separate Class

This does not mean every public venue should be empty.

An auction hall without any public context can feel false. But public context is not the same as named-character continuity.

The useful taxonomy is:

Visible Class	Function	Risk
Named character	speaks, reacts, blocks space, carries power, handles prop	can pull focus or invent action
Public crowd texture	establishes public venue pressure	can become too detailed or active
Physical anchor	stabilizes space, light, material, position	safest continuity carrier

A good public crowd texture is low-detail, static, non-speaking, and subordinate to the current subject. It should establish public pressure without becoming a new dramatic actor.

That distinction matters. "A few low-detail seated guests in shallow background" is a scene-class anchor. "Blurred family members behind the speaker" is often a named-character leak.

Prompt Difference

A weak scene prompt says:

Warm luxury living room, tense atmosphere,
family members in the background.

A production-safe version says:

Daytime living-room sofa zone.
Cool natural light enters from the window side.
Matte floor remains non-reflective.
Dark wood tea table fixed lower center in front of the rear sofa.
Only the current speaker is clear.
No other named family member is visible unless this shot uses them as listener or action partner.

The second version is more controlled because it is smaller.

It does not describe the whole world. It describes the current zone.

It does not ask the model to remember a previous render. It repeats the physical facts that must survive this local request.

Lab Judgment

The more I work with AI video, the less I trust mood as a continuity device.

Mood is useful for taste. It is not enough for production memory.

Scene memory should be compiled from physical facts: time, light, floor, furniture, material, prop state, zone, and allowed visibility. The public prompt should receive only the subset that belongs to the current shot.

The goal is not to describe the world beautifully.

The goal is to keep the world from mutating when the model regenerates it.

For AI video production, continuity is not created by longer prompts. It is created by smaller, stricter, better-scoped prompts.

A scene is too large.

A zone is closer.

A shot function is closer still.