Reference image limits, and why they differ so much

Every model that accepts a reference image publishes its own limits, and they differ enough that one asset library cannot feed all of them. The ranges below are what four vendors state on their own API pages. As of 2026-09-12.

Inclusion rule. Models whose vendor publishes reference-image limits on an API or documentation page. A model accepting references without publishing limits is not given a row. Order. Alphabetical by model name.

1Build the library to the tightest limit

The intersection of these ranges is narrower than any single one: a JPEG or PNG, comfortably under 20 MB, between 300 and 5760 pixels on a side. An asset built to that specification feeds every model here without conversion.

Building to one vendor's maximum and converting later is how a season ends up with two libraries and a conversion step that quietly changes the images.

2Transparency and exotic formats are where it breaks

One vendor excludes PNG with transparency specifically. HEIC is accepted by two and unmentioned by the rest. A library holding either will work until the day a shot is regenerated on a different model, which is exactly when nobody is looking for a format problem.

3Regenerating with a different reference is a different request

Swapping a reference for a better one mid-season changes the output even when the prompt is unchanged. That is fine when it is deliberate and a continuity error when it is not, so the reference used for each delivered shot belongs in the record.

4Aspect ratio limits are the quiet one

Ratio ranges are published as well as pixel ranges, and a tall crop of a character can fall outside them while still looking correct. Where a reference is rejected without a useful message, the ratio is the first thing to check.

5A reference is not a style guide

Everything visible in the image is a candidate for being copied: the background, the lighting, the crop, the compression artefacts. A reference shot against a busy room supplies the room as well as the subject, and the model has no way to know which part was meant.

The practical rule is to strip the image down to the thing being referenced. It looks unhelpfully plain and produces far more predictable output than a handsome frame does.

6Where the limits are usually discovered

Rarely at the start. A library built from screenshots passes for weeks, then a shot fails on one model because a PNG carries transparency or a crop fell outside a ratio range, and the failure message says nothing useful.

Checking a sample asset against the published limits once, at library-building time, removes a class of problem that otherwise surfaces at the worst moment.

7Where these figures come from

• MiniMax video API reference
• Seedance reference limits
• Vidu reference-to-video documentation
• Wan 3.0 API reference

A definition kept for this register: the term, and what vendors state about it, each figure read on 2026-09-12. The tools themselves are on the register. Related: Character consistency, Storyboard script.

8One pipeline's continuity

Serial work like “Reference image” depends on continuity; SceneMixer's guides describe how it keeps it:

The same face across an entire series is the design target and what the reference-sheet architecture delivers for the cast in your asset library.

SceneMixer, AI Video Generator With Consistent Characters — “How consistent is it, really?” (read 2026-10-13)

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)