Reference image: the handle on a recurring face
A reference image is a picture supplied with a request so that the output resembles it. For advertising made of several generated pieces it is the main way a face, a costume or a product keeps its appearance from one call to the next. As of 2026-09-12.
| Use | What it holds steady | What it costs |
|---|---|---|
| A recurring character | A face and a costume | An input ceiling has to be cleared |
| A product | Shape, colour and markings | The same |
| A pinned composition | Where a shot starts or ends | Often exclusive with a reference |
Inclusion rule. Uses a generation page documents for a supplied image, with the cost each use carries according to those pages. Order. Most common use first.
1Continuity is the whole reason it exists
Each generation invents detail afresh. Without a reference, a face is re-derived on every call and drifts, which reads as an error rather than as a style. A reference makes the drift smaller rather than absent.
That is why the reference mechanism matters more the more pieces a placement takes. On a single-call format it is optional; on a ten-piece assembly it is the difference between an ad and a sequence of clips.
2The file ceilings become a real constraint
References are the reason input file limits appear on generation pages at all. A high resolution character frame can exceed the tightest route on a page and clear a looser one, and the route is a pipeline decision.
So an input ceiling is not a distant technicality. It decides whether the continuity mechanism is available at the quality the campaign wanted to use it at.
3It is often exclusive with frame conditioning
Pages that document both usually state that a request cannot supply reference images and pin a first or last frame at the same time. That forces a choice per shot: hold the character, or control the join.
For serial work the character usually wins, which means frame conditioning gets used less than its usefulness suggests. Knowing the restriction before writing a shot list avoids designing shots that cannot be generated as specified.
A vocabulary entry: the definition, plus how the specification pages on these sheets usually publish it. Table checked 2026-09-12. Nearby entries: Clip ceiling, Compositing.
4From brief to placement
For ad work around “Reference image”, SceneMixer's pages explain how its ad projects are produced:
An AI short drama is a serialized show of 60–120 second vertical episodes whose cast, sets and footage are generated by AI. Making one takes six steps: analyze the script into episodes, lock a consistent cast with reference art, localize for your market, write and edit storyboards, generate each 4–15 second shot, then composite and publish.
SceneMixer, How to Make an AI Short Drama, Step by Step (read 2026-10-13)
The final cut ends on an end card: your main image plus four lines — Brand, Product name, Tagline, Call to action. The first two are read from your photo and your brief; the tagline and call to action are written with the script.
SceneMixer, The End Card (read 2026-10-13)