One to seven images per call, and nothing kept between calls
The reference documents one to seven reference images per call, supplied so the model produces a consistent subject, and documents no reusable character or element held between calls. Input limits are published as well, which is how a developer's page differs from a product page. As of 2026-09-12.
| Point | What the vendor's page states | How this file reads it |
|---|---|---|
| The sentence | One to seven reference images per call, for a consistent subject | Documented in a developer reference |
| What it names | The request, the images and the stated input limits | Everything the caller controls, and nothing it does not |
| What it leaves out | Any object surviving the call | Recorded as not documented, which is what the page supports |
Inclusion rule. Three points per reading: the sentence as the vendor published it, the thing that sentence names, and what it declines to settle. Nothing here is inferred from an interface or a demo. Order. From the published sentence, to what it names, to what it leaves open.
1The cast lives in the caller's filing system
Sending the same images on every request will produce a subject that looks the same, so the continuity is real. What holds it is a folder on the caller's side, along with whatever discipline keeps that folder correct across forty episodes and more than one person.
That is a legitimate arrangement and a common one. It is also the opposite trade from a stored cast, where the vendor owns the artefact and also owns what happens to it when the underlying model is replaced.
2Published input limits are a different kind of generosity
A minimum pixel size and an allowed aspect-ratio range are the sort of numbers a product page never carries and an integration cannot proceed without. They also make this the one reading in the column where a reader can tell in advance whether their existing stills qualify.
The limits say nothing about continuity, and that is the point of recording them separately. Precision in one place is not evidence about another.
3Where this sits in the column
| Tool | What the vendor publishes on the cast |
|---|---|
| AI Drama Generator | Consistent characters listed among the outputs, with no mechanism named |
| AI Dramas | Each character saved with a name, a face and a voice style, then called by name in every script |
| Anijam | Locking characters' faces across every episode stated as an action the tool performs |
| Arcloop | Locking a face, outfit, personality and visual identity listed as a step, with no object named |
| character.app | A recurring Cast and World described as objects the series is built on |
| Crano AI | The same identity stated to hold across episodes, with stable appearance and emotional presence |
| DramaSo | Consistent characters listed among the outputs, with no mechanism named |
| Dramily | Character dynamics, scenes and props turned into reusable assets |
| Genra | Cast cards holding a face and a voice, stated to carry across every scene |
| Google Flow | Ingredients, the same images reused for a character and key objects from one clip to the next |
| invideo AI | Project memory holds character sheets and locations |
| Kling AI | Elements carry appearance and voice between generations |
| LTX Studio | Elements are reusable across projects |
| Luma Agents | Characters, products and brand elements defined once and preserved across named operations |
| Medeo | Consistency attributed to the video model the product runs on |
| Novi AI | A character created in the product, whose appearance the system locks for the whole project |
| OpenArt | Leads stated to stay the same across every episode, face, wardrobe and chemistry |
| Pexo | Hyperframe, named as what locks a face, wardrobe and voice across every episode |
| PopShort AI | Characters, scenes and props stored as reusable production assets, from Episode 1 to Episode 50 |
| SceneMixer | Reference sheet per character, crops injected into every segment |
| Showrunner Studio | Sets, characters and canon stated to persist across the whole series |
| StoryShort | Casting consistent characters listed as a step, with no mechanism named |
| Topview | Characters and scenes stated to be built before video generation, to help continuity |
| TrendShort | Locking identity, wardrobe, face and emotional range listed as a step |
| Vidu | One to seven reference images supplied per call |
Inclusion rule. Every entry on this file appears in the column below, including the ones whose vendor says nothing on this step. The marked row is the entry this page is about. Order. Alphabetical by tool name.
4The wording itself
- Cross-episode castOne to seven reference images per call, used so the model generates a video with consistent subjectsper call rather than per character
- Cross-episode castNo reusable character or element stored between calls is documentedwhat the documentation does not describe
- Stated limits on the inputs and the outputReference images must be at least 128x128, with an aspect ratio inside 1:4 to 4:1; durations run 1 to 16 seconds depending on model
5Sources
Read from the reference-to-video API reference at platform.vidu.com, checked 2026-09-12. The whole column is on cross-episode cast and this vendor on every column is on Vidu; where the line falls between documented, claimed and absent is under how read. Other readings of the same column: PopShort AI, Showrunner Studio.