How to Build a Reusable Location Library So Interiors Stay Consistent Across 70 Episodes

The apartment in episode 6 has a grey sofa against a window on the left. In episode 31 the sofa is charcoal, the window has moved to the right, and there is a bookshelf that has never existed before. Nobody approved a redress. No note asked for a change. The apartment drifted because two different operators described it in two different sentences eleven weeks apart, and the model did exactly what it was told each time. Environment drift is the quietest failure in long form AI vertical drama, and it is the one that most reliably makes a delivered series feel cheap.

Character consistency gets attention because faces are the thing audiences notice consciously. Environments are noticed unconsciously, which is worse, because the audience registers that something is wrong without being able to name it. On a 70 episode series with eight to twelve recurring interiors, environment drift is not a risk to be managed. It is a certainty unless a library exists. The method below sets out how to build one that survives a full series and carries into a sequel.

1. The Problem Is Re-Description, Not Generation Quality

Every time an interior is described in words rather than referenced as an asset, a new version of it is created. The model is not remembering the apartment from episode 6. It is reading a sentence and producing something that satisfies that sentence. Two sentences that a human reads as identical, a small modern living room with a grey sofa and a large window, will produce two visibly different rooms, because the sentence underdetermines the room by an enormous margin. The failure is in the workflow, not in the tool.

This is why prompt discipline alone does not solve environment consistency. A locked text block gets closer, and teams that enforce one see less drift than teams that do not, but text will never fully specify a space. The distance between the sofa and the window, the ceiling height, the direction the light falls, the colour temperature of the practicals: these are continuous variables that a paragraph cannot pin down. What pins them down is an image that the generation is conditioned on.

So the shift required is small to state and significant in practice. Interiors stop being descriptions carried in a prompt library and become assets carried in an asset library. Each recurring space is captured once, approved once, and referenced from then on. Operators stop writing rooms and start calling them.

2. Decide What Counts as a Location Before You Build Anything

A location is a space that appears in more than one scene and that an audience is expected to recognise as the same place. That definition excludes a great deal of what a script might call a location. A corridor a character walks down once is not a library asset. A car interior used in fourteen scenes is. Building the library around the wrong unit is the most common way this work becomes expensive without becoming useful.

Run the script and count. Every interior, with a scene count against it. Anything at four scenes or above goes into the library without discussion. Anything at two or three gets a judgement call based on whether the space carries story weight, because a room that appears twice at a story turning point matters more than a room that appears five times as background. Anything appearing once is generated ad hoc and forgotten. On a typical 70 episode series this produces a library of roughly ten to fourteen interiors, which is a manageable build.

Do the count before generation starts, not after. A library assembled retroactively from frames already delivered inherits whatever drift already occurred, and the team then has to decide which version of the apartment is canonical while thirty episodes carry the other one. Retrofitting is possible and it is always more expensive than building in pre production.

3. Capture Each Interior as a Plate Set, Not a Paragraph

For each library location, generate and approve a set of environment plates covering the angles the series will actually need. In practice that is four to six plates per space: a wide establishing view, two opposing mid angles that support shot reverse shot coverage, a detail or insert angle, and where relevant a view through the main light source. These are generated once, reviewed as a set, and locked. They become the conditioning images for every scene set in that space.

Generate the plates empty. No characters, no action, no expression. An environment plate that contains a person will push that person into downstream generations and will bias the framing toward wherever they were standing. The plate is the room. Characters arrive later, conditioned separately, placed against it.

Review the plate set as a set, not plate by plate. The question is not whether each image is attractive. It is whether the six images describe one coherent room, with consistent geometry between the angles and consistent light direction across them. A plate set where the window is on the left in one angle and the left in the reverse angle is internally contradictory, and it will produce an interior that feels subtly wrong for the length of the series. Reject and regenerate at this stage. It is the cheapest correction point that will ever exist.

4. Write the Location Card

Each location gets a card that travels with the plates. The card is short and it is authoritative. It carries the canonical name of the space, the plate filenames and what angle each covers, the locked descriptive block that accompanies the plates in a prompt, the light direction and colour temperature, and a list of the fixed set dressing that must appear. Anything not on the card is not in the room, which gives an operator a clear rule rather than a judgement call.

The card also carries the exclusions, and these matter more than teams expect. If the apartment has no plants, the card says no plants. If the office has no visible branding, the card says so. Models fill empty space with plausible objects, and the objects they choose vary between generations. Naming the absences is how a space stays stable across eleven weeks and four operators.

Keep the card to a single screen. A location card that runs to three pages will not be read by an operator working through a batch at pace, and an unread card is functionally identical to no card. The discipline is to encode the small number of things that actually drive recognition and to leave the rest to the plates.

5. Name and Store So the Library Survives the Series

Naming is the part that gets skipped and the part that determines whether the library is reusable in eighteen months. Every plate and card follows one convention: series code, location code, angle code, version. LOC_APT_WIDE_v2 tells an operator everything needed and sorts correctly in any file browser. Descriptive filenames written by whoever generated the plate will not survive contact with a second operator, and they certainly will not survive a sequel commissioned a year later by a different team.

Storage follows standard digital asset management practice rather than anything specific to AI production. One location per folder, cards alongside plates, superseded versions retained rather than deleted, and a single index that lists every location with its code and scene count. The index is the thing a new operator reads on day one, and it is the thing a platform receives at delivery if the series is archived for sequel use.

Version rather than overwrite. When a location changes for story reasons, a fire, a redress, a time jump, that is a new version of the location and not a replacement. Episodes already generated against version one remain correct. Episodes from the change point forward reference version two. Overwriting destroys the ability to regenerate a shot from episode 12 six weeks later, which is a request that arrives on most series at least once.

6. Handle Time of Day and Season as Variants, Not New Locations

The same apartment at night is not a second apartment. Treating it as one doubles the library, doubles the review burden, and introduces the risk that the two versions drift apart from each other. Instead, each location carries variants: day, night, and whatever additional states the script requires. A variant inherits the geometry and the dressing from the base location and changes only the lighting condition.

Generate variants from the base plates rather than from scratch. Conditioning the night version on the approved day plate holds the room geometry while changing the light, which is exactly the desired behaviour. Generating the night version independently from the text description reintroduces the original problem in a form that is harder to spot, because reviewers compare day frames to day frames and night frames to night frames and rarely compare across.

Cap the variant count. Two or three per location covers almost every vertical drama requirement. A location with six variants is usually a sign that lighting states are being used to carry story beats that should be carried by performance or coverage, and it is worth a conversation with the direction side before the library absorbs the complexity.

7. Govern Additions So the Library Does Not Sprawl

Libraries fail by growing, not by staying small. An operator hits a scene in a space that is not in the library, generates something workable, and moves on. Three weeks later there are nineteen unofficial interiors in circulation and no one knows which are canonical. The governance rule is simple: only a named owner adds to the library, additions go through the same plate set and card process, and anything generated ad hoc is explicitly marked as one off so it is never mistaken for an asset.

Give the library one owner for the duration of the series. On most Axis AI Studios productions that sits with the production coordinator rather than with a generation operator, because the role is custodial and administrative rather than creative. The owner runs the index, approves additions, and is the person an operator asks when a space is ambiguous. One owner is enough. Two owners produce two conventions.

Weekly is the right review cadence. The owner checks what was generated that week against the index, catches any space that was improvised, and either promotes it to a library asset or confirms it as a one off. Five minutes a week prevents the situation where a series wraps with an asset set that nobody can hand to the next production.

8. Audit the Library Against Delivered Episodes

The library is a control, and a control that is never tested is a belief rather than a control. Once every ten episodes, pull the same location from the oldest delivered episode and the newest, put them side by side, and look at geometry, dressing and light direction. This is a five minute check and it is the only reliable way to catch slow drift, because drift that accumulates at one percent per episode is invisible between adjacent episodes and obvious across thirty.

Extend the check to continuity editing logic rather than stopping at appearance. Does the geography hold across a cut? If a character exits frame right in the living room and enters the kitchen, is the kitchen where the plate set says it is relative to the living room? Spatial relationships between locations are not captured by individual plates and are the next layer of drift after appearance, particularly on series where characters move between rooms within a scene.

Record the audit result in the index. Two lines per check: what was compared and whether it held. When a platform asks at delivery how environment consistency was controlled, the index with its audit history is the answer, and it is a considerably stronger answer than an assurance that the team was careful.

Axis AI Studios Perspective

A location library is infrastructure, and infrastructure gets built before production or it does not get built. The cost is roughly two days in pre production for a 70 episode series: counting spaces, generating and approving plate sets, writing cards, setting the index. The cost of not building it is distributed across the whole series as retakes, reviewer time spent adjudicating which version of a room is correct, and a delivered product that reads as inconsistent to a platform buyer who cannot articulate why.

Axis AI Studios is an AI native vertical drama production studio based in the Netherlands. Environment libraries are standard on every series produced, alongside character reference sets and prompt libraries, because all three address the same underlying property of the format: a long series generated in batches over weeks by more than one person drifts unless assets are locked and governed. That is a production management position rather than a technical preference.

The secondary return is the one platforms tend to value most. A library built to this standard is the asset that makes a sequel cheap. A second series in the same world starts with approved interiors, a named index and an audit history rather than starting again, and the economics of a franchise depend heavily on whether that work was done the first time.

To discuss production standards for a specific series or a commissioning slate, write to business@axisaistudios.com.

FAQ

How long does building a location library actually take?

For a 70 episode series with ten to fourteen recurring interiors, expect around two working days in pre production. That covers the scene count, generating four to six plates per location, the review pass that rejects and regenerates incoherent plate sets, writing the cards and setting up the index. Series with heavier location counts or with significant variant requirements will run longer. The work is front loaded by design, because a library assembled during production inherits whatever drift has already occurred.

What happens if a location needs to change partway through the series?

It becomes a new version rather than a replacement. The original plates and card stay in place so that any episode generated against them can be regenerated later, and the new version carries a clear effective point, typically an episode number. Operators working forward from that point reference the new version. This matters more than it seems, because requests to regenerate an earlier shot arrive on most long series, and a library that has overwritten its own history cannot service them.

Can the library be reused if the production tool changes?

The plates and cards remain valid because they are images and specifications rather than tool artefacts. What may need revalidation is how a new model responds to the plates as conditioning input, so the sensible step after any tool change is to regenerate two or three frames in the highest volume location and compare them against delivered episodes before committing a batch. Where a plate no longer conditions reliably, regenerate that plate from the existing approved version rather than from the text description.

Further Reading

For the visual specification layer that sits above individual locations and governs how the whole series looks, the guide to building a vertical drama style guide covers palette, lighting and framing standards across 70 episodes.

For the equivalent discipline applied to people rather than spaces, and for how an asset set carries from one series into a franchise, the walkthrough on building an AI character asset library covers structure, storage and reuse.

For the tracking layer that records what changed in each space and when, the guide to building a vertical drama continuity bible covers environment state alongside character and narrative state.