EpisodeReel

Serial work: what each tool documents about continuity

Reference images per call are not a stored cast

The documented interface takes one to seven reference images per call so the model generates a consistent subject. No reusable character or element stored between calls is documented. Inputs must be at least 128x128 within a stated aspect ratio range. As of 2026-09-12.

Where consistency is stored under each mechanismA documented interface that accepts references per call keeps nothing between calls, so the file of reference images held by the caller is what makes episode twelve match episode one. A product with a stored cast moves that responsibility, and the failure modes move with it.References per callA stored cast objectWhat the interface takesOne to seven reference images on eachrequestA saved character, by referenceWhat is documented aspersistingNothing between callsThe character itselfWhere the images liveWith the caller, re-sent every timeWith the productWhat the input rules apply toEvery call: at least 128x128, inside thestated ratio rangeWhatever the product accepted when thecharacter was madeTwo mechanisms, two different production jobs
Fig. 1 Neither mechanism is better. They put the versioning discipline in different places, and only one of them is visible to the production.
What the documented interface does with a character, read from the API reference. Recorded 2026-09-12.
AxisWhat Vidu documentsDetail
What is acceptedOne to seven reference images per callPer request, not per project
What persistsNothing documented between callsNo stored character object
Input limitsAt least 128x128, aspect ratio inside a stated rangeApplies to every call

Inclusion rule. Recorded from the developer documentation rather than a marketing page. Absence of a stored object is recorded as not documented, which is what the pages support. Order. From what goes in, through what is kept, to what constrains it.

1The consistency lives in the caller, not the product

Handing the same images to every call will produce a subject that looks the same, and the thing keeping the series consistent is the caller's own file of reference images. That is a real solution and it puts the storage, the versioning and the discipline on the production.

A product with a saved cast makes the opposite trade: consistency is the vendor's problem, and so is what happens to it when the underlying model changes.

2Why the distinction survives into episode twelve

A stored object can be edited once and applied everywhere. A per-call reference has to be re-sent correctly every time, which is fine for ten calls and a source of drift across a season assembled over weeks by more than one person.

Neither approach is recorded here as better. What is recorded is which one each tool documents, because the production work each implies is different from the first episode.

3Input limits are part of the cast question

Minimum dimensions and an allowed aspect ratio range decide what may be used as a reference at all. A production whose character stills are crops from a wider frame can fall outside them without anything in the interface appearing to refuse the work.

They are recorded in the same row as the reference mechanism for that reason: they are not a separate technical footnote, they are the rules the cast has to satisfy.

  • Cross-episode cast
    One to seven reference images per call, used so the model generates a video with consistent subjectsper call rather than per characterVidu, reference-to-video API reference / recorded 2026-09-12
  • Cross-episode cast
    No reusable character or element stored between calls is documentedwhat the documentation does not describeVidu, reference-to-video API reference / recorded 2026-09-12
  • Stated limits on the inputs and the output
    Reference images must be at least 128x128, with an aspect ratio inside 1:4 to 4:1; durations run 1 to 16 seconds depending on modelVidu, reference-to-video API reference / recorded 2026-09-12

4The words this register uses

Terms with a page of their own, each recording what vendors publish about it, include segment and shot, reference image, character consistency, storyboard script, scene plate, novel to video, AI short drama and vertical micro-drama.

5Continuity in practice

Serial work like “Per call, not per cast” depends on continuity; SceneMixer's guides describe how it keeps it:

That is SceneMixer's core design: characters live in a persistent asset library with locked reference art, and every segment that features a character is generated against those references — same face in episode 1 and episode 50.

SceneMixer, The AI Micro Drama Production Pipeline — “Can AI keep the same actors across 50 episodes?” (read 2026-10-13)

SceneMixer separates the script language from the production language: the output (episode outlines, storyboards, on-screen text) follows your interface language, and target-region localization adapts character names, casting, and scenery for Western, East Asian, Southeast Asian, or Latin American audiences.

SceneMixer, Turn a Novel into a Video Series with AI — “Can I turn a Chinese web novel into an English-language drama?” (read 2026-10-13)

6Sources

Each cell above traces back to a vendor page named in the register, read 2026-09-12. Where the boundary sits between documented, claimed and absent is spelled out under how read. Also here: Season length, A rate or a plan, Where it was published.