EpisodeReel

Serial work: what each tool documents about continuity

Four ways to keep a face the same across shots

Four techniques are in common use: a reference sheet supplied to every generation, reference images passed per call, a reusable character object the platform stores, and a trained adapter. They differ in who keeps the state and in what breaks when it drifts. As of 2026-09-12.

Where a character appearance is kept, in each of the four mechanismsFour columns feeding one generation call. Per-call references and a reference sheet both keep the appearance in your own files and send images into every call. A stored character object keeps it on the platform and sends only a name. A trained adapter keeps it in the model weights and sends nothing, because by then it is the model. Under the call, what is left to inspect afterwards: exactly what you sent, the sheet itself, nothing you can open, and the training set.Per-call referencesYour filesImages, on every callExactly what you sentReference sheetYour filesThe same sheet, every callThe sheet itselfStored characterobjectThe platformA nameNothing you can openTrained adapterModel weightsNothing: by then it is themodelThe training setOne generation callWhat is left to inspect afterwards
Fig. 1 The mechanisms differ in one thing before they differ in anything else: where the appearance is kept between generations. That choice decides what there is to look at on the day a shot comes back wrong.
Four mechanisms, and what each one costs. Recorded 2026-09-12.
MechanismWhat it needs
Per-call reference imagesAttaching them every time
Persistent reference sheetBuilding it once, up front
Seed reuseThe platform exposing a seed
Text description onlyNothing, and it does not hold

Inclusion rule. Mechanisms that appear in published material. Approaches requiring model training are a different question and are not listed. Order. Alphabetical by mechanism.

1A reference sheet is the cheapest thing that works

A character sheet is a small set of images that fix an appearance: a front headshot, a three-quarter angle, a full body, and often one or two poses. Supplied to every generation, it gives the model something concrete to match rather than leaving it to reinvent a face from a text description each time.

It costs nothing beyond the images themselves and it is fully inspectable: you can look at what the model was given. Its weakness is discipline. Nothing enforces that every generation receives the sheet, and a shot produced without it is indistinguishable at the moment of generation and obvious in the cut.

The single highest-value habit in this whole area is deciding the sheet before the shot list exists, rather than assembling it after the first drift appears.

Three views of the same woman on a plain grey backdrop: a head and shoulders portrait, a three-quarter view from the waist up, and a full-length standing shot, in the same jacket and with the same short dark bob in all three.
Fig. 2 A reference sheet in its plainest form: one character, three views, a plain backdrop and even light. Nothing in it is a performance. It exists so that every later generation has something concrete to match instead of a sentence to interpret.Generated for this page with an image model on 2026-09-16. Not output from any tool held on the register, and not a frame from any production.

2Per-call references put the burden on the pipeline

Some systems take reference images on the call itself, several at a time, and use them to keep subjects consistent within that generation. This is the same idea as a sheet with a different owner: the platform stores nothing, so the reference set is a production asset somebody has to version and attach correctly every single time.

That is more work and it has a real advantage: nothing can drift invisibly, because there is no hidden state to drift. What you sent is what was used, and it is in your own records.

3A stored character object moves the state to the platform

Defining a character once as a reusable object, then referring to it, is less work per shot and harder to inspect. When output stops matching, the question becomes what the platform thinks the character is, and that is not something you can open and look at.

Scope matters here too. An object reusable only inside one production is a different asset from one reusable across productions, and for a studio with a recurring cast that difference decides whether the cast is a project file or a company one.

4Trained adapters are the heavy option

Training a small adapter on many images of a character encodes the appearance into the model itself. It is the strongest form of lock and the most expensive: it needs a corpus of images, a training run, and a model you are allowed to adapt, which in practice means open weights.

The rule of thumb people arrive at is scale. For a handful of controlled shots, references are enough. For hundreds of shots with very little drift tolerance, the heavier stack starts to pay for itself, and by then the production is large enough to afford it.

5Seeds are a smaller lever than they look

Fixing the random seed across related shots removes one source of variation and is free, so it is worth doing where the platform exposes it. It does not hold a face together on its own: a fixed seed with a vague description still produces a different person when the prompt changes.

Treat it as a tidiness measure that makes comparisons meaningful, not as a consistency technique.

6What actually causes drift, in rough order

A missing reference on one generation. A prompt that describes the character differently from the sheet. A model version change midway through a production. A shot at an unusual angle or distance the sheet does not cover. And regeneration weeks later against a newer context.

Four of those five are production discipline rather than model capability, which is worth sitting with: most drift is fixed by deciding things earlier rather than by choosing a different tool.

A woman at a rain-streaked window at night, lit by a warm lamp from the side, with the same blunt fringe and rust-red jacket as the reference sheet.The same woman standing full length on an empty railway platform under flat grey daylight, in the same jacket and boots as the reference sheet.The same woman seated at a wooden table beside a window in daylight, hands on the table, with the same face and fringe as the reference sheet.
Fig. 3 Three shots with the sheet above attached to each generation. The setups have nothing in common: night interior close, overcast exterior wide, daytime interior medium. The face, the fringe, the jacket and the boots survive all three.Generated for this page on 2026-09-16, each with the sheet above supplied as a reference image. Not output from any tool held on the register.
A woman at a rain-streaked window at night in a rust-red jacket, with a side-parted dark bob and a different face from the reference sheet.A woman on a railway platform in a rust-red jacket, longer hair and a different face from the reference sheet, photographed closer than the wide shot it was asked for.A woman seated at a table by a window in a rust-red jacket, with a different face and no fringe, unlike the reference sheet.
Fig. 4 The same three setups, the same wording for the character, and no sheet attached. The jacket comes back every time because it is in the sentence. The face does not, because a sentence has no face in it.Generated for this page on 2026-09-16 from the character description alone, with no reference image. Not output from any tool held on the register.

7Where this points, and where it does not

If the priority is a cast that survives forty episodes with the least ceremony, the approaches worth reading first are the ones that attach appearance to a named thing rather than to a habit: SceneMixer generates a reference sheet per character and injects it into every segment, which makes the attachment automatic while leaving the images inspectable; Kling AI and LTX Studio define reusable elements, with LTX Studio's scoped across productions.

If the priority is auditability, the per-call approach that Vidu documents is the one where nothing is hidden, at the cost of somebody attaching the right images every time. None of this is a quality judgement, and what each vendor has actually published is on the register.

8Where the register holds these tools

The techniques above are general craft. Which tools document support for them, and in what words, is held in the register, each cell carrying the page it was read from and the date.

  • SceneMixer — documents a full-body reference sheet generated per character and injected into every segment
  • Kling AI — defines characters as elements that carry appearance and voice
  • LTX Studio — makes elements reusable across productions rather than inside one
  • Vidu — takes one to seven reference images on each call with nothing stored between them

Craft notes, not a register entry. No line here is attributed to a tool, and none should be read as a claim about one. The sourced material is on the register. Related: Continuity log, The six stages.