By CJ Simon-Gabriel |
Beyond generation: our first step towards iterative editing
Every generative AI product has a retry button. You've used it. You press it because something in the output is wrong. Maybe a sentence, a pair of hands poorly captured, a second of audio that just doesn't hit right — and the only way to fix that one thing is to dismiss the whole thing and roll the dice again.
We've grown used to this. We shouldn't have.
The retry button reveals something about how most generative models work today. They can produce a clip or an image, but they generally cannot revise it. They cannot change one second or detail while leaving the rest untouched. Each generation starts over, with no relationship to the one before. This is expensive: in time, compute and creative momentum.
This is not how craft works. Craft is iterative. A music composer fine-tunes their piece. A film editor returns to a cut. A sound designer re-edits to the second where the footstep lands wrong, and changes only that. The defining act of making something well is the process of iteration.
Generative models have been celebrated for going forward. For producing. For generating. The benchmark has been: can the model make the thing? And the answer, increasingly, is yes, it can. What almost no audio model can do is the harder part: stay with what it has made, understand it, and edit parts of it.
That is the gap Mirelo SFX 1.6 begins closing.
And it matters more for audio than people realize. "Sound is 50 percent of the movie going experience," said George Lucas. And it's the 50 percent that carries most of the emotion. Getting it right isn't optional. Neither are the tools to do it.
What it means for a model to iterate
Mirelo SFX 1.6 is our first model that can iterate on its own output. It enables audio inpainting, which unlocks a form of local editing: regenerating parts of an audio clip and blending it back seamlessly into its surroundings. It's a meaningful first step towards full iterative editing; and the step we're taking today.
Mirelo SFX 1.6 comes with four new capabilities
All available today across Mirelo Studio, our Adobe Premiere and DaVinci Resolve plugins and our API.
Long generations – up to 60 seconds
Need a full minute of rain building to thunder, a low engine rev or an extended crowd swell to go along with your video? Our new model lets you generate sound tracks that are six times longer than before. Generate it all in one pass. No stitching, no seams.
Extension
An existing clip (generated or recorded) is great but too brief? Enter the length you want, and the model picks up exactly where your sound leaves off.
Looping
Generate soundscapes that naturally end where they begin. Crowd cheers, engine hums, city ambience, and more. The perfect ambient loop from a simple text prompt – or from existing audio.
Inpainting
Clean a sound clip by removing unwanted sounds. Select the glitchy section of an almost perfect audio clip and we'll fix it without touching the rest. The seconds left and right stay as they were; the new section fits in smoothly.
What's ahead
Iteration is the real benchmark — the one that craft has always demanded. Mirelo SFX 1.6 is our first model built around that conviction. Local regeneration and extension open a new world of possibilities to bring more control to creators. But we're just getting started…
Sound on.
