Which AI video model is best for ads?

There is no single best. Veo 3.1 leads on native audio and lip sync for on-camera speech, Seedance 2.0 on reference-driven product fidelity, native audio and single-take UGC up to 15 seconds, Seedance 2.5 on single-take length — a measured 30 seconds with no cut in it, with native audio — and Kling 3.0 on face-reference identity, the same person across renders, plus image-to-video. The right pick depends on the shot, and the ranking shifts with each release — Hermoso routes every job to the model best suited for it.

How long can a single AI-generated take be?

Thirty seconds, on Seedance 2.5 — we asked for thirty and measured 30.08 in one continuous shot with native audio and no hard cuts. Fifteen seconds was the ceiling before it, which is why longer ads have been stitched out of several clips. One pass means no seam to hide: no grade shift, no room tone dropping between beats. It renders at 480p and 720p, so a spot that has to finish at 1080p is still a job for a shorter take.

Which model has the best lip sync?

In our experience Veo 3.1, because it generates speech natively in the same pass as the footage, so the lips and audio come from one process. The uncanny lip-flap comes from dubbing separately generated TTS over a rendered mouth — reserve that only for cuts where no mouth is visible.

Which model keeps my real product looking right?

Seedance 2.0 holds product identity well when it's fed a real product photo as a reference and the object is composited into the scene. Text prompts describe categories; reference images pin identity, which is why grounding on a genuine photo beats describing the product in words.

Which model keeps the same actor across a campaign?

Kling 3.0. Feed it a face reference and it holds that person's identity across renders, so a recurring creator can front a whole series of ads instead of a new face appearing in every spot.

Why do AI videos struggle with thrown or dropped objects?

Ballistic motion — an object in free flight under gravity — is where every current model's physics still visibly breaks, regardless of which one you use. Choreograph around it: keep products held, turned, set down, or poured so a hand carries the object and the model never has to solve gravity on its own.

Do I have to choose the model myself on Hermoso?

No. Hermoso routes each render to the model best suited for that shot and re-tunes the routing as new versions ship, so you never have to track a changelog or pick a model. You describe the ad; the routing handles the rest.

hermoso

Blog

Seedance 2.5: 30 seconds of ad in one unbroken take

We measured it the day it shipped: a 30-second ask came back at 30.08 seconds with zero cuts and native audio — and it animated a real photograph of a person, which the previous version refuses outright.

By Mara Vivanco·August 7, 2026·6 min read

Most model releases move a benchmark. This one removes a workaround.

Until now, a video ad longer than fifteen seconds was not one video. It was several, stitched. Every generator we route caps a single clip somewhere around fifteen seconds, so a thirty-second spot meant rendering it in pieces and hiding the joins — matching the light, keeping the same person looking like the same person, opening each new clip mid-thought so the seam did not announce itself. A lot of our prompt engineering exists for no reason other than to make those joins invisible.

Seedance 2.5 renders thirty seconds in one take. We asked for thirty and measured what came back: 30.08 seconds, zero hard cuts, native audio. Not a stitch that hides well — a continuous take with nothing to hide.

What we actually measured

Everything below is from renders we ran on the day it shipped, not from a spec sheet.

Thirty seconds, one take. A 30s request returned 30.08s. Scene detection found zero hard cuts across the whole clip.
It animates a real photograph of a real person. We fed it an ordinary portrait and got a photoreal clip of that person back. The previous version refuses this outright — it rejects a photoreal human face before rendering anything.
Native synchronised audio, including speech, on by default.
480p and 720p. There is no 1080p on this model. Full-HD work still belongs on Seedance 2.0.
Up to 30 image references, plus video and audio references, addressed by name inside the prompt.

The face thing matters more than the length

The thirty seconds is the headline, but the quieter change is the one that alters what you can make.

Putting a specific person in an AI ad has meant choosing between two bad options. Describe them in words and the model casts someone who matches the description — photoreal, but a subtly different person in every clip, which is fatal the moment your ad has more than one shot. Or lock the identity to a stylised reference and hold the same face throughout, at the cost of it looking drawn rather than filmed.

A real photograph on Seedance 2.5 gives you both at once: the same person, and photoreal. That is not a better trade-off. It is the trade-off going away.

Where 2.0 still wins

We have not retired it, and we are not going to pretend the new one is strictly better.

Seedance 2.0 is the only tier that renders 1080p and 4K. If you are making something that needs to hold up at full resolution, that is still where it goes. 2.5 tops out at 720p, and we would rather say so plainly than let you find out in the export.

So the honest shape is: 2.5 for length, for a real face, and for anything that benefits from one continuous take. 2.0 when resolution is the constraint.

What this changes about writing an ad

A thirty-second single take is not a longer version of a fifteen-second one. It is a different kind of shot, and it wants to be written like one.

A stitched spot is a sequence of beats, each doing one job, joined at the cuts. A continuous take is a move: something is set up, something happens without interruption, and it lands. If you write thirty seconds of beats and hand them to a model that will not cut, you get thirty seconds of drift. If you write one motion with an arc, you get the shot the model is actually good at.

It is live in Hermoso now

Seedance 2.5 is the default video model in Hermoso as of today, and reachable by name over MCP and the CLI if you want to pin it. Ads up to thirty seconds now render as a single take rather than being assembled from parts.

If you want the wider comparison — how it sits against Veo 3.1 and Kling 3.0, and which shot suits which engine — that is in our model comparison.