Seedance vs Kling vs Veo vs Hailuo: which job each one fits in an ad

Choosing between Seedance, Kling, Veo and Hailuo gets easier when you ask what job each one does in an ad, not which one wins. This page compares them that way, from our own work with several models. Every model fact comes from the maker’s own documentation. There is no winner here and no prices. In our own ads in autumn 2026, Seedance did the b-roll; that is our history, not a recommendation.

The jobs in a video ad

A finished ad asks a model for several different things. “Which model wins?” skips that. “Which model fits this job?” does not. The full lineup sits on the Turntake models page, and the whole chain around it is in how to make AI UGC ads.

Source: our own production rules, October 2026.

1. B-roll
2. A product moment
3. A screen for the real form

What each maker says

The next table lists only what the makers’ pages state, with one source per row, checked October 2026. It is not ranked.

Source: ByteDance Seed, Google, Kling AI and MiniMax documentation, checked October 2026.

Two cautions. These pages tell you what a model accepts and returns, not how well it follows your prompt. And they move: the Seedance 2.5 post describes a rollout on Jimeng AI, Doubao Pro and other platforms, with API access to follow. MiniMax lists Hailuo 2.3 among its legacy models and H3 as the newer one. Read the linked page again before you plan a batch.

Still or video?

“Seedance vs Nano Banana” is not a contest, because they do different jobs. Google’s documentation describes Nano Banana as image generation and editing through conversation. Seedance moves a picture. We make the still with an image model, approve it, and only then ask a video model to animate it. The still rules are in Nano Banana for ads.

1. 01The stillAn image model
2. 02You approve itYou
3. 03The motionA video model

Talking heads and voice are other jobs

A talking head is a different job from b-roll, and the maker sets a rule for faces: ByteDance Seed’s launch post for Seedance 2.0 says real human portraits used as references need identity verification or prior legal authorization. In our tests in October 2026, through one API provider, Seedance 2.5 refused our photoreal faces, an observation from one provider rather than a verdict on the model. So we give the face to a talking-head model and keep the video model on b-roll. The method is in AI UGC talking heads, and the voice, a job of its own, is in ElevenLabs for ad voiceovers.

How we test a model against our own spec

We don’t run a bake-off once and keep the result forever. We write the spec first and test every candidate against it.

1. Write the spec as pass or fail lines: what must be in the frame, and what must not. For a portrait: front-facing, no phone in hand.
2. Run each candidate on the same approved still with the same prompt shape.
3. Judge on a numbered contact sheet, with frames from across the clip rather than only the first one.
4. Keep the model that obeys the spec and stays photoreal. Don’t write up the others as bad. They failed this spec on that day.
5. Run the test again when the models change.

The same habit applied to talking heads. In July and August 2026 we tried several routes and kept the one where lip-sync, identity and reliability held. Routes built on long multi-segment character references let the identity drift between segments, and “continue from the previous clip” mixed up scenes. That is what we tried, not a ranking.

Prompt rules we carry into every test

These come from image-to-video work, mostly on Seedance, since September 2026. We haven’t run every rule on every model, so treat them as a starting spec. Kling’s own image-to-video guide gives related advice: keep the language simple, and know that a description far from the image can cause a camera cut or transition.

• One move, one action. One simple camera move, one subtle action, and a closing line like “keep the scene exactly as in the reference image”. Consistency comes from the locked start frame; if your provider also offers a seed, keep it fixed.
• One place per generation. In our tests, a change of place inside one run changed the figure’s face at the cut. Split the tail into two runs, and start run B on the last frame of run A. Write dialogue as timestamped beats, one line per beat, or the model may treat it as a pile of lines and improvise.
• No generated text. In our tests, generated text turned into gibberish. We state that every sheet of paper is blank, and we burn numbers on afterwards.

More detail on these is in how to use Seedance for ads.

Cost without prices

We don’t list prices. Each maker publishes its own, and they change. What we can share is how we keep a bill predictable.

• Resolution follows purpose: standard for b-roll that sits under captions, higher only for shots with text or numbers. The finished master is 1080×1920.
• Don’t generate sound for b-roll that will be silent anyway.
• Library first. A new ad generates only the beats the library lacks.
• Every job has a hard cap and a cost log, written when the job is sent. Test one clip, get it approved, then run the batch.

Today we run those caps and logs ourselves, as described in how to use Claude for ads. In Turntake, Cost shown before you run Coming is what we’re building so you see the cost in credits before anything runs, and Models in Claude Coming is the plan for reaching these models through one connection. Early access: we set it up with you. Coming: we’re building it.

Which model and voice questions need their own guide?

Use the detailed guide for the task in front of you. Each owns one question, so you can work through a specific production decision and return to this overview for the wider workflow.

• Face-detected refusals
• Consistent characters
• Voice pronunciation
• Music bed planning
• Word-level caption timing
• Kling shot planning
• Veo shots and sound
• Hailuo image-to-video
• Dubbed lip-sync
• Start and end frames
• Legible generated text
• Hand and object details
• Voiceover word budgets
• Voice auditions
• Voiceover pacing

Who wrote this We made the ads first, for SmartSolar in Denmark, Sweden and Ireland. Then we built Turntake. Where Turntake comes from