Grok Imagine 1.5 just dropped, and your characters can finally keep their voice

Half a million views in a day. That is how fast word of the Grok Imagine 1.5 update travelled, and for once the noise is about something practical.

The headline feature, shown off by creator @dvorahfr: pre-recorded voices you can pin to a character. Pick a voice once, and every animation featuring that character sounds like the same person. Voice drift has been one of the quietly broken things in AI video, the detail that gives away a stitched-together film in the first ten seconds. If this holds, it stops being a post-production job.

The update is also unusually easy to try. Runway confirmed on its official account that Grok Imagine Video 1.5 landed on their platform on the 1st, available now rather than coming soon. A model update you can test somewhere other than home on day one is still rare enough to be worth noting.

Why voice consistency is the expensive problem

If you have only made single shots, this feature reads as a nice-to-have. If you have made anything with a character who appears twice, you know it is not. The standard workaround today is to take the voice out of the model entirely: generate silent video, record or synthesise the dialogue separately in a dedicated voice tool, then lip-sync it back on. That works, and it is three extra steps, two extra subscriptions, and a sync pass on every clip.

Pinning a voice at the model level collapses that. It also changes what you can attempt: episodic content with a recurring character, an ad series with the same spokesperson, anything where a viewer meets the same face twice and expects continuity. Those formats have been technically possible and practically miserable. The interesting question is not whether the voice matches within one generation, it is whether it survives a week, a different prompt, and a different scene, because that is where consistency features usually break.

Against what people use now

The honest comparison is not model versus model, it is integrated versus assembled. ElevenLabs and similar tools give you far more control over performance, direction and emotion than any pinned preset will, and if voice is the star of your piece you will keep using them. What an in-model voice buys you is speed and coherence on everything else: the fifteen-second social cut, the variant test, the shot you need today.

Early testing is enthusiastic and, as ever, selective. @minchoi called the update wild to an audience of nearly half a million views. Creator @hedo_ist ran 1.5 against Seedance 2.5 on a deliberately awkward test, forcing close-range interpretation of faces shot from a distance, and rated the result excellent. Treat individual tests as anecdotes rather than benchmarks; the pattern across several of them is the signal.

What to test first, and what nobody has answered

Generate the same character in three unrelated scenes on different days and listen for drift. Push it through an emotional range, because presets often hold timbre and lose delivery the moment a line needs anger or a whisper. Check what happens with two characters in one scene, which is where voice assignment tends to get confused. And find out whether a pinned voice survives a prompt rewrite or is quietly tied to the phrasing that created it.

What has not been said publicly: whether you can supply your own recorded voice or only choose from a library, how many voices you can hold, and whether any of this is available through the API rather than the interface. Those three answers decide whether this is a production feature or a demo feature.

For now the practical takeaway is simple enough. Locked character voices plus same-day Runway availability moves 1.5 out of the curiosity pile and into the daily-driver conversation. If your pipeline already lives in Runway, you can answer most of these questions yourself this afternoon.

Sources: Runway's official announcement, @dvorahfr's voice demo, @minchoi, @hedo_ist. 1 to 3 August 2026.