The 18 video models, grouped by what they're actually for

Vendor names group these models by who made them. That's almost never the question. What you actually want to know is which one fits the shot, and that comes down to four numbers.

What actually differentiates them?

Four settings, and every model in the picker exposes all four before you generate.

Duration. The range runs from Veo 3.1's locked eight seconds to Seedance 2.5's thirty. Most sit at fifteen.

Resolution. From 480p at the bottom to 4K on four models, with one offering 1440p and one offering 2K.

Reference capacity. This varies more than anything else — from one reference on Hailuo 2.3 and Kling 3.0 Turbo to fifty on Seedance 2.5. Veo 3.1 Lite has no references section at all.

Audio control. Six models give you a toggle. Twelve have sound always on with no way to disable it.

The vendor matters far less than any of those.

Twelve of eighteen models have audio you cannot switch off. That's the setting nobody checks and the one that breaks a workflow.

The long-take group

Two models, and only one of them genuinely holds a long shot.

Seedance 2.5 is the outlier on two axes at once. Thirty seconds in a single pass is twice what any other model manages, and fifty references is more than three times the next highest.

Vidu Q3-Turbo reaches sixteen seconds and its description names what it's for — first and last frame. It's a keyframing model rather than a long-take one, and the duration is a by-product of interpolating between two anchors.

Both have an audio toggle, which matters if you've cast a voice separately.

The reference-heavy group

Five models where reference capacity is the point.

Where this falls short. Reference capacity tells you what a model will accept, not what it will render well. Fifty references in one generation is technically available and almost certainly a bad idea.

The reference-composed group

Three models whose descriptions all name composition from references, at lower ceilings.

Kling 3.0 Omni's description is "richest reference control — frames, video, elements," and the seven-reference ceiling makes sense once you read it as a claim about types rather than volume. It's the breadth of what a reference can be, not how many you can attach.

HappyHorse 1.1 has the widest aspect ratio range of any model — nine options, including 9:21 and 5:4 that nothing else offers. Its description names native audio and lip-sync in one pass, which puts it alongside the dialogue models too.

Wan 2.7 starts at two seconds, the shortest minimum on the roster.

The Veo group

Four models, and the differences between them are narrower than the naming suggests.

Two things stand out.

Veo 3.1 Fast gives up nothing on capability. Same eight-second lock, same three resolution tiers including 4K, same three references. The only difference the picker exposes is the description — "faster Veo, same sound and scene extension." So it's a speed tier rather than a capability tier.

Veo 3.1 Lite is the one that actually costs you things. No references section at all, no 4K, and a 4–8 second slider rather than a lock. Which makes it more flexible on duration and less capable everywhere else.

Gemini Omni Flash has no duration or resolution controls whatsoever. Its panel goes references, prompt, aspect ratio, sound — and nothing else. It's the only model on the roster with no length or quality decision to make, which its description explains: "any input in, sound-matched video out."

Veo 3.1 Fast is a speed tier. Veo 3.1 Lite is a capability tier. The naming suggests the opposite.

The draft group

Four models built for iteration rather than delivery.

Two genuine oddities here.

LTX-2.3 Fast offers four resolution tiers including 1440p and 4K, which is more than most delivery models. Its description is "rapid drafts from text or a frame," and its two-reference ceiling fits that — but the resolution range doesn't.

Hailuo 2.3 has no aspect ratio control at all in Create mode. Its panel runs length, then quality, with nothing between. Every other model with a framing decision exposes it; this one doesn't offer one.

The dialogue group

Two models whose descriptions name synchronised audio specifically.

Both have audio always on, which follows from what they're for — a lip-sync model with sound disabled isn't doing its job.

Grok Imagine 1.5 starts at one second, which alongside Vidu Q3-Turbo is the shortest minimum available. Useful for a single beat or an insert rather than a scene.

For a face delivering a script, though, neither of these is usually the answer. Talking Actors derives duration from the script and gives you emotion control at specific words, which no prompt matches.

What should you actually pick?

Work from the constraint, not the vendor.

Need more than sixteen seconds in one pass? Seedance 2.5. It's the only option.

Need 4K? Seedance 2.0, Veo 3.1, Veo 3.1 Fast or LTX-2.3 Fast. Four models, and Seedance 2.5 isn't among them.

Need to turn the audio off? Seedance 2.5, 2.0, 2 Fast, 2 Mini, Vidu Q3-Turbo, or Kling 3.0 Omni. Six of eighteen.

Need many references? Seedance 2.5 at fifty, then the 15-reference Seedance tier, then MiniMax H3 and HappyHorse 1.1 at nine.

Need an unusual aspect ratio? HappyHorse 1.1 with nine options, then Seedance and Grok families with seven.

Iterating cheaply? Seedance 2 Fast or 2 Mini keep the 15-reference ceiling while dropping to 480/720.

All figures read from Hexcoded's Creative Studio picker in Video → Create mode in September 2026. Model specifications change with versions, and availability differs by mode and by plan. Verify in the picker before relying on any of this for a commercial delivery.

More on model capability, access and rights in Models.