What native audio actually gets you, and where it fails

It's listed as a feature and it's actually a workflow decision. The question isn't whether native audio is good — it's whether the sound is a decision on your project.

What does native audio actually do?

Generates the audio in the same pass as the picture, rather than as a separate step you align afterwards.

Hexcoded's own model documentation names the mechanism. Seedance 2.5 uses an Audio-Visual Diffusion Transformer architecture, processing visual frames and acoustic signals within the same generation pass — where traditional video generation creates silent footage first and overlays external audio tracks afterward. Visual movement and sound dynamics are calculated simultaneously, so footsteps, environmental ambience, surface impacts and spoken dialogue align with visible kinematics without a post-synchronisation step.

That sounds like a convenience and it's structurally more than one. When you sync separately, you're matching two things that were produced independently, and the match is something you maintain. When the model produces both together, there's nothing to match.

The clearest place this shows is dialogue in close-up, where a small sync error is immediately visible and a matched take simply doesn't have one.

What do you actually get?

Four things, and only the first is what the feature name suggests.

Where this falls short. The third row is the one people underestimate. Getting ambience automatically is convenient until you want a specific soundscape, at which point automatic ambience is something you're working against rather than with.

Is it one setting, or per model?

Both, and the distinction catches people.

Creative Studio presents a native sound toggle as a global control across all four video sub-modes. But per-model behaviour overrides it. Seedance 2.5 gives you a toggle you can switch off. Wan 2.7 has sound always on with no toggle at all — and the picker describes Kling 3.0 Turbo as "fast cinematic clips, audio always on."

So "does this model do native audio" has three answers rather than two: toggleable, always on, or absent. Which means the question to ask isn't whether a model supports sound — it's whether you can turn it off.

That matters more than it sounds. If you've planned to sync separately and the model generates audio you can't disable, you're mixing against a bed you didn't ask for.

The question isn't whether a model does native audio. It's whether you can turn it off.

What does it cost you?

Control, in three specific places.

Voice casting. A natively generated voice is what the model produced for that character on that take. If the voice is a decision — a specific age, accent, register — native audio doesn't give you that lever. Talking Actors is the tool where voice becomes a casting choice, because the actor carries both the face and the voice.

Consistency across a series. Hexcoded's own Seedance 2.5 documentation names the mechanism for the visual case: latent re-encoding compounds small variations when subject descriptors are altered across chained generations. The same logic applies to sound. For a recurring character across many episodes, that's the same drift problem in a different channel.

The mix. Dialogue, ambience and effects arriving as one layer means you can't balance them independently. For a finished piece that's a real limitation; for a draft it's irrelevant.

Native audio removes a match problem and adds a drift problem. Which one you'd rather have depends on how long the series runs.

That's the honest framing. For a single piece, the match problem is the bigger one and native audio wins. Across a season, consistency becomes the harder problem and separate audio starts to look better.

When should you sync separately?

Four cases, and they're all about the sound being a decision rather than an output.

When you've cast a specific voice, whether a performer or a synthetic voice you've chosen deliberately. When a character recurs across many episodes and consistency matters more than per-shot convenience. When the soundscape is doing narrative work and needs building rather than accepting. And when you're delivering to a client who may want the sound changed after the picture is approved.

That last one is the practical trap. A picture-locked edit with baked-in audio is hard to revise, and "can we change the read?" is an ordinary note.

There's a constraint worth knowing here too. Where a human creator appears, Hexcoded's terms permit light edits only — you may not re-voice a face outside the platform. So "sync separately" means using a model where audio is off, not re-voicing a delivered take externally.

How do you use audio as an input?

The part most people miss, because it works the other way round from what you'd expect.

Audio isn't only an output. Creative Studio's video engine accepts up to 10 audio clips as references inside its 50-asset pool, and Hexcoded's own documentation describes their role as rhythmic cues, acoustic tone and synchronisation context.

Which means you can supply a track and have generation take pacing from it, rather than generating and then cutting to music. For anything rhythm-led — a montage, a beat-matched sequence, an ASMR piece — that's the difference between editing to a track and generating to one.

Where this falls short. Ten audio references is a ceiling, not a target, and the reference pool is shared. Audio references come out of the same 50 assets as your images and actors.

How should you decide, per project?

Four questions. The first yes settles it.

Model capabilities described here are published specifications and observed picker behaviour current as of the publication date, not measured comparisons. Audio behaviour differs by model and changes with versions. Verify in the picker before relying on this for a delivery.

More on model capability, access and rights in Models.