Prompts get longer when you supply a reference image: 2,942 characters versus 1,313

By Alpha Lay · 2026-09-26

"Once you have a reference image, the prompt can be simpler, right?"

That is the natural assumption — the look is given, the model just has to animate it. After pulling prompt lengths across 1,274 cases, we found the opposite.

Where the data comes from

Each case records a pc field: the character count of that case's prompt text (counted in the original language, with no conversion across languages). All 1,274 have a value. Median 1,563 characters, mean 2,666, minimum 4, maximum 24,828.

The distribution is wide to begin with:

Over a third (39.0%) run past 2,000 characters.

Finding 1: The modes that need reference material run 2.2–2.5x longer

First/last-frame prompts are 2.5x the length of text-to-video; reference-to-video is 2.2x.

Once you think it through, it stops being counter-intuitive: reference material fixes what someone looks like, not what they do. When the character's appearance is locked, action, camera, timing and transitions all still have to be specified in words. Text-to-video can get away with "a girl walks down the street" and let the model improvise — it does not even have to lock a face.

Put another way: a reference image removes "appearance" from the prompt but raises the bar on "behaviour." Total length goes up, not down.

Finding 2: A 2.6x spread across genres

Action & physics (2,261) runs 2.57x animals & creatures (880).

The ordering makes sense: action content has to specify direction of force, contact points, chain reactions — things adjectives cannot cover, so they get written out line by line. Animal clips are usually one subject doing one thing and need a few sentences.

Worth correcting a common impression: product ads (1,429) are shorter than short-drama dialogue (1,553). Commercial content does not mean complicated prompts — it is often a single shot, a single action, a clean background.

Finding 3: Very short prompts come with shorter clips

The shortest band's median duration is 13.7s, clearly below the 15.2s of every other band.

That lines up with our duration piece: 15.2s is the model's default output. With a very short prompt, the model is more likely to wrap up early and land below the default. Writing in more detail makes it more likely to fill the default length.

What you can do with this

If you use reference images or first/last frames: do not assume the image replaces the writing. By the data, these are the longest prompts of any mode — around 3,000 characters at the median. Put the effort into action and camera description, because appearance is already handled.

If you teach prompt writing: here is a directly usable rule — ask first what this piece relies on for locking things down. A reference image locks the subject, so spend everything on behaviour. Without one, you must also describe the subject, but you can be more general about it.

If you build prompt templates or tools: a 2.6x spread across genres means no single universal template can be right at the length level. Action content needs a much longer skeleton than landscape content.

Method and limitations

• Character count is not information content. Mixed Chinese and English, embedded JSON, parameter notes all affect pc, and the same information is usually shorter in Chinese than English. This piece makes relative comparisons only and does not interpret absolute values.
• pc provenance is not uniform: some prompts are verbatim from the publisher, others come via third-party indexes and may be truncated or completed. We did not verify every original.
• The 930 "mode unverified" cases are included (median 1,398), but since their mode is unknown, they cannot support the claim about which mode is longer — that claim rests on the 344 mode-verified cases.
• A data-hygiene note: the minimum pc is 4 characters, which is far more likely a parsing error than a genuine 4-character prompt. We did not remove it, because no consistent rule for removal exists — but readers should know this band contains noise.
• All figures use the 1,274 index records.

Reproduce it

Case detail views show each record's prompt and its length: https://cases.aishifu.shop/