5.3 Holding the World Together: Outfits, Locations, Props

A consistent face in an inconsistent world reads as a generated face. The audience's judgment is not "is this person the same person" in isolation; it is "is this the same life." A kitchen that changes between posts breaks that judgment as effectively as a jawline that changes.

Three asset libraries fix it, all built the same way: generate once, validate, store, reuse.

Outfits: a library, not a decision

The face has an adapter. Outfits have a reference image each, and the reason is structural: an identity adapter trained on images of a person also learns what they were wearing. Left alone, it will re-render that clothing in new scenes. Chapter 4.4's wardrobe variety reduces the effect; an outfit reference eliminates the need to fight it.

What an outfit asset contains:

One clean image of the persona wearing it, front view, neutral background, even light
The complete prompt fragment that produces it
The season and occasion it belongs to (from 2.3)
The accessory set: shoes, bag, jewellery, nails

The target size: twelve to twenty outfits for a first year. Fewer and the feed repeats visibly; more and you cannot hold them consistent anyway.

Why this is cheaper than it looks. A wardrobe of fifteen outfits, each applicable to five locations, gives seventy-five wardrobe-scene combinations from twenty generations of work. The combinatorics are the whole point, and they are the practical return on 2.3's style section.

The rule for changing outfits: adding an outfit is a production cost; replacing one is a continuity event. The audience registers a disappearance as much as an arrival. Retire an outfit deliberately, and note it in the register from 2.4.

Locations: the highest-return asset

A location can be generated once and reused for years. It has no identity to hold, no expression, and no drift. If you build only one library in this chapter, build this one.

How to generate a reusable location:

Generate a clean plate — the empty space, no person, at the highest resolution you use. This is the canonical version.
Generate the same space from three to five additional angles, as separate clean plates. This is what prevents every post being the same corner of the same room.
Name them consistently (home-kitchen-a, home-kitchen-b, home-kitchen-counter).
Store them as references, not as images you re-generate from scratch.

The angle set that avoids repetition: a wide establishing view, a medium view with a surface in the foreground, a detail shot (the counter, the window), and one view from the doorway looking in. Four angles per location is enough for hundreds of posts.

How many locations: five to eight is the working range for one persona. Chapter 2.3's twenty places is the prompt closet; five to eight of them get built as reusable sets, and the rest appear once or twice within a scene.

The consistency test: generate the same location twice, a week apart, without the reference. Compare. If the window is in a different wall, the room's identity is not held yet — and that inconsistency is more visible to an audience than a slight facial difference, because people are extremely good at spatial memory.

The closed-world rule: the smaller the persona's world, the cheaper it is to keep consistent. A persona who lives in one apartment, walks one street, and visits three places is dramatically cheaper to produce than one who travels constantly. The trade is between variety of setting and cost of consistency, and the cheaper side is usually the better-performing side — a recognizable recurring set is what makes an account feel like a place rather than a feed.

Props: the cheapest specificity

Props convert a generic render into a specific world. A laptop, a mug, a plant, a specific chair, a book, a pair of shoes by the door.

Why they matter: a scene with no objects is a scene with no evidence of a life. Objects are also the easiest thing to hold consistent, because an object has no — it either matches the reference or it does not.

Building the prop set: from the in 2.3, pick the ten objects that appear most often in your content plan. Generate a clean reference for each, alone on a neutral background, and reuse them across every location.

Props have a second function: they carry continuity. A plant that grows, a mug that travels, a jacket that appears across seasons. Recurring objects are how an audience builds a mental model of the world, and they are the cheapest narrative device in this medium (2.4).

The reference tree

Where all of this lives, using 3.6's folder structure:

00-bible/          character bible, canon, timeline register
10-references/
  face/            the 3×3 sheet, versioned with the checkpoint
  body/            full-height front and back
  outfits/         12–20 wardrobe assets
  locations/       5–8 locations × 4 angles, as clean plates
  props/           ~10 object references
20-datasets/       training images (4.4)
30-lora/           adapters + hashes (4.5)
40-output/         generated frames
99-archive/        retired versions

Two rules about this tree. References are never overwritten — a new version gets a new name and the old one moves to archive. And every output records which references produced it, which is the provenance manifest from 11.5 and the skill you need when a turns out wrong.

Quarterly asset audit

The libraries accumulate and rot quietly. One pass per quarter, in the review slot from 2.6, keeps them usable.

Prune unused references. Any outfit or prop that has not appeared in three months is either wrong for the current direction or forgotten. Retire it explicitly and move it to archive, so the active set stays small enough to remember.

Check location drift. Regenerate one image against each location's reference and compare with the plate. Locations drift as models and versions change, and a location that has drifted produces frames that no longer match anything else in the feed.

Rebalance the combination usage. Look at the last quarter's published frames. If two of eight locations account for most of the content, the feed is narrower than you think, and the weekly combination sheet is not being applied.

Record what you added and retired, with the date. This feeds the from 2.4 and answers the question "when did the old jacket disappear" — which is the kind of question an audience asks and an operator cannot answer from memory.

The audit takes an hour and it is the difference between a reference library and a folder of old images.

Coordinating without a shot list

At ten posts a week, coordination is memory. At thirty, it is not, and the failure is specific: duplicate outfits, a location appearing three times in a week, the same prop in every frame.

Two lightweight mechanisms are enough.

A weekly combination sheet. Before generating, list which outfit, location, and light each post uses. Thirty seconds of work that prevents the visible repetition. This is a step in the production slot from 2.6.

A rolling exclusion. Nothing that appeared in the last ten posts appears in the next ten. Simple to apply and it enforces variety mechanically, which is exactly what you want when you are generating in volume and cannot remember Thursday.

When to break consistency on purpose

Deliberate incoherence is a tool, and it should be deliberate.

Seasonal change. The whole kit's grade shifts warm or cool, some outfits rotate out. Scheduled in 2.4's map.

Planned events. A trip, a move, an event. The world changes visibly, once, on purpose. This is where new locations enter the library.

Appearance change. Hair, styling, fitness — the visible evolution from 2.4. One or two a year.

A deliberate continuity break as content. The behind-the-scenes format from 2.6 can show the retired outfit and the new one side by side. It converts a continuity cost into content.

The distinguishing feature of all four: they are recorded in advance. Unrecorded variation is drift; recorded variation is a story.

Where this goes wrong

Only building the face reference. Outfit and location inconsistency is more visible than small facial differences.

One location, one angle. Hundreds of posts from the same corner.

Regenerating locations per post. Spatial memory is acute; the kitchen will move.

A world that is too big for you to hold. The most common cause of inconsistency at scale, and the opposite of the closed-world rule.

Props appearing without references. Every new object is a new inconsistency risk.

Overwriting a reference. You have lost the ability to explain or reproduce what you published.

Coordination by memory. It works at ten posts a week and fails at thirty.

The gate

Do you have a 3×3 face sheet, a body sheet, and 12–20 outfit assets?
Do you have 5–8 locations with four angles each, stored as clean plates?
Do you have ~10 prop references?
Is your world small enough that you can hold all of it consistently?
Is there a weekly combination sheet, and a rolling exclusion you actually apply?
Are references versioned, never overwritten, and linked to outputs?

What you leave with

Three asset libraries — outfits, locations, props — built from clean references and reused rather than regenerated, the combinatorics that make a small library produce a year of content, the closed-world rule and its real trade, a reference tree that keeps every layer versioned and traceable, two lightweight coordination mechanisms that hold at thirty posts a week, and the four legitimate reasons to break consistency deliberately.