How many references can a holiday ad take? Five video models compared

Seedance 2.5 lists the most reference slots per pass at 30 images, 10 video clips and 10 audio clips, and Wan 3.0 lists up to 20 reference assets. On Sume, Gemini Omni Flash 1.1 takes up to 10 reference images and 3 reference videos of up to 3 seconds. Luma's Ray3.2 uses up to 16 keyframes, and Sume does not run it.

The numbers are vendor limits, not Sume limits, apart from the Omni row. Read what the Sume catalog reports for the model you call. This post helps you decide how large a reference kit you should plan, since a holiday ad usually needs a product, a logo, a person and a style.

The limits side by side

Every row below is from the vendor's own page, read 2026-10-05, except the Sume row for Omni Flash 1.1, which comes from the Sume Video Router docs. Luma Ray3.2 is not in the Sume catalog, so it is a comparison only.

What the limits mean for an ad kit

A holiday kit is small. One product photo, one logo, one lifestyle reference, one voice sample, one music cue. That is five items, below every model's limit. The limit that bites is not the count. It is what the reference types let you do.

Audio references matter if the ad needs a consistent voice or a beat. The Sume docs say Seedance 2.x, Wan 3.0, MiniMax H3 and H3 Max accept audio and video references, while Gemini Omni Flash 1.1 accepts video references but not audio. If the voice has to match across a series, choose a model that accepts audio references, or use Sume's text-to-speech for the voice and add it in assembly.

Video references let you carry a motion. Omni's 3-second, up to 3 clips limit is enough for one move, such as a pour or a spin, but not for a whole sequence.

Length and price

Seedance 2.5 and Wan 3.0 both go to 30 seconds on Sume, and Omni Flash 1.1 stops at 10. If your ad is a 30-second story with several references, the first two are the candidates. For a 10-second hook with a single product shot, Omni is enough and often cheaper.

The per-second rates are in the Sume pricing tables: Wan 3.0 is $0.0625, $0.125 and $0.25 per second at 480p, 720p and 1080p, Omni 1.1 is $0.125 at 720p, and MiniMax H3 is $0.075 at 768p. A 10-second clip is therefore $1.25, $1.25 and $0.75.

Cautions

Two cautions. Vendor limits change, and a vendor page can describe a feature that an API provider has not exposed yet. Seedance 2.5's own post (read 2026-10-05) said the API was coming via BytePlus ModelArk at that time, while Sume lists seedance-2.5 in its catalog, so the Sume docs are the authority for what you can call. And a higher limit is not a better ad. A kit of three clean references beats a kit of thirty mixed ones.

Build the smallest kit that works

Start from the job, not the limit. If the ad needs a product to stay on-model, send two or three clean product photos from different angles and nothing else. More views of the same item help a model hold its shape, while unrelated images tend to blend into the result.

If the ad needs a logo, send it as a flat image on a plain background and say in the prompt where it goes. If it needs a person, send one face reference and one outfit reference, and do not mix several people. If it needs a style, a single frame from a previous approved ad works better than a mood board.

Test the kit on one cheap clip. Render a 5-second draft at the lowest resolution with the full kit, and look at whether each reference is visible. Then drop the reference that did nothing and render again. A kit that survives that test is the one to reuse across a campaign, and it is cheap to store: keep the list of reference URLs next to the prompt.

Reuse the kit across a series

For a series of ads that must look related, keep the same kit and change only the prompt. For a series that must look different, change the style reference and keep the product references. Either way, record the kit with the job so a later edit can reuse it. Reference URLs must be reachable when the job runs, so host them somewhere stable for the whole season rather than on a temporary link.