Can AI regenerate one section of a cue when the edit changes?
October 20, 2026
Yes. Stable Audio 3.0 regenerates a marked region of an existing cue and leaves the rest of the file untouched, using start and end times in seconds. Several non-contiguous regions can be rebuilt in one pass, so a recut that moves four moments does not require a new cue.
Key takeaways
Region regeneration rebuilds a marked span of an existing cue and preserves everything outside it.
Both mask parameters accept lists, so several scattered fixes go in a single generation.
A cue the production generated itself has no rights holder to approach when the picture changes.
Stable Audio 3.0 does not watch the cut, so a swell landing on a new frame is still a manual job.
What a picture change asks of a cue
Picture changes arrive after the music is working, which is the whole difficulty. A scene loses four seconds, a montage gains a shot, a note from the studio collapses two beats into one, and the cue that was sitting correctly under all of it now resolves in the wrong place. Music editors conform what they can. Past a certain amount of movement, the cue has to change rather than slide.
For a music supervisor the operative question is who else has to be involved in that change. A licensed master is someone else's recording, and whether it can be re-edited, re-timed or extended depends on the terms of that particular license, which may require going back to the rights holder before anything happens in the DAW. Reporting may also need revisiting once timings move.
A cue the production generated carries none of that. The file belongs to the production, the revision is an editorial decision, and the only constraint is what the tool can actually do to it.
Who you go back to when the cue has to change
Ownership is the part that makes a generated cue revisable on the production's own schedule. Under the Stability AI Community License, outputs of Stable Audio 3.0 belong to the user, who may use them at their discretion in compliance with applicable law and the acceptable use policy. Commercial use under that license is free for organizations generating under $1M in annual revenue, and organizations above that figure contact Stability AI about licensing.
Nothing in that position changes when a section gets rebuilt. A regenerated region is another output, owned on the same terms as the original.
How Stable Audio 3.0 regenerates one region
The mechanic is called inpainting, and it comes down to two numbers.
Marking the region by timecode
Stable Audio 3.0 takes an existing audio file, a start time and an end time, regenerates only the span between them, and preserves everything outside it. The parameters are literal seconds, inpaint_mask_start_seconds and inpaint_mask_end_seconds, exposed on the command line as --inpaint-start and --inpaint-end and in the Gradio interface as mask start and mask end, per the Stable Audio 3.0 repository. An editor who loses four seconds at 00:32 regenerates 32 to 36.
Region size behaves counterintuitively, and Stability's own Stable Audio 3.0 prompting guide is direct about it. Masking only a few seconds gives the model so much surrounding context that the result can come back close to the original. Start with a large mask and tighten it on later passes.
Several fixes in one pass, and a scene that grew
Notes rarely arrive one at a time. Both mask parameters accept lists, so passing starts of 4.0 and 16.0 with ends of 8.0 and 20.0 rebuilds two disconnected windows in a single generation. Four fixes scattered across a sequence resolve together.
Scenes that grow use the same mechanic. Setting the mask start to the length of the source file and asking for a longer duration carries the existing material forward into the extension. A 20-second clip masked from 20s to 100s comes back as 100 seconds.
| What changed in the cut | What region regeneration does | What still sits with the team |
|---|---|---|
| Four seconds lifted at 00:32 | Rebuilds 32 to 36 and leaves the rest byte-identical | Checking the join against picture |
| Notes scattered across one sequence | Rebuilds every marked region in a single pass | Deciding which regions are actually wrong |
| Scene grew past the end of the cue | Extends the cue from its existing material | Keeping the extension in the same arrangement |
| A hit needs to land on a new frame | Nothing. Duration is exact to the second, not the frame | Placing the hit in the DAW |
| The cue is a licensed recording | Nothing. The file was not generated | A revision may require going back to the rights holder under that license's terms |
Region-level editing sits alongside two other modes, text-to-audio and audio-to-audio, covered in more depth in our explainer on generating stems and inpainting regions of a track.
What region regeneration will not fix
Stable Audio 3.0 does not watch the cut. Generation runs from a text prompt and a duration with no video input, so the model has no knowledge of where the door slams. An editor describes the scene rather than uploading it.
A musical moment you liked does not survive a regeneration. Masking 32 to 36 rebuilds those four seconds from scratch, and the phrase that was there is gone, so the moment to protect is the one to leave outside the mask.
Three further limits are worth planning around. Prompts have to be plausible against the surrounding material, because a request that sits far from the existing track tends to be ignored. Output is instrumental, with no singing or lyrics, per the Stable Audio 3.0 model overview. Cue length tops out at 380 seconds on Medium and 120 seconds on the Small models, which bounds how far a continuation can run.
Productions generating temp cues rather than revising them should start with generating an original temp cue instead of licensing one.
Frequently asked questions about recutting a generated cue
Does the whole cue have to be regenerated when the cut changes? No. Stable Audio 3.0 rebuilds only the span between the start and end times you set and leaves the rest of the file intact. A four-second lift at 00:32 means regenerating 32 to 36, not re-rolling a 94-second cue and losing everything that already worked.
Can several sections be fixed in one pass? Yes. Both mask parameters accept lists, so starts of 4.0 and 16.0 with ends of 8.0 and 20.0 regenerate two separate windows in one generation. Non-contiguous editing in a single call is documented in the Stable Audio 3 repository.
What happens if the scene gets longer instead of shorter? Continuation covers it. Setting the mask start to the current length of the file and choosing a longer duration extends the cue from the material already there. A 20-second clip masked from 20s to 100s returns 100 seconds.
Can a specific musical moment survive a regeneration? No. Everything inside the mask is generated fresh, so a phrase you want to keep has to sit outside the boundaries. Tightening the mask around the problem rather than the passage is the practical approach.
How long a section can be replaced? Any span below the model's ceiling, which is 380 seconds on Medium and 120 seconds on the Small models. Short masks tend to return audio close to the original, because the surrounding context dominates, so Stability recommends masking large and reducing on later passes.
Does regenerating a section change who owns the cue? No. Under the Stability AI Community License, outputs belong to the user, and a regenerated region is an output on the same terms as the original. Whether a cue can be registered as a copyrighted work is a separate question, and criteria for copyrightability remain under active debate.
Can a cue the production did not generate be regenerated this way? The model will accept any seed audio, including a recording made in-house. What a production is permitted to do to a licensed master is set by that license, and re-editing or extending it may require approval from the rights holder.
Can this run inside the cutting room on unreleased picture? Yes, with the open-weight models. Small runs on CPU with no GPU and Medium runs on a CUDA GPU, so an assistant editor can regenerate against unreleased footage on a machine inside the facility. No audio and no scene description leaves the building.
Last updated October 20, 2026
