Stable Audio 3 Inpainting: Repair Two Bars Without Rebuilding the Cue

An instrumental cue works except for one awkward fill. Regenerating the entire piece risks losing the opening, the groove, and the ending you already approved. Stable Audio 3 inpainting offers a more focused question: can you replace the problem region while keeping the surrounding cue useful?

Download the audio inpainting region map · Develop a cue in Music Lab

Feature information checked September 17, 2026. The eight-bar cue is an original worked timing example, not a reported Stable Audio generation. The hero is an AI-generated editorial illustration.

Choose inpainting when the problem is local

Stability’s guide to stems and inpainting distinguishes generating parts from replacing a selected region. Its official Stable Audio 3 prompting documentation describes inpainting with a start and end region. Controls differ between hosted interfaces and the documented workflows, so verify that the environment you use actually exposes inpainting.

Use it when a specific passage is wrong: a fill distracts from narration, an ending lands too heavily, or a transition changes character. If the entire cue has the wrong tempo feel or instrumentation, a local repair may only conceal a broader brief problem.

Do not confuse audio-to-audio transformation with region replacement. A workflow that uses the whole source as conditioning may reinterpret much more than the two bars you wanted to fix. Confirm the mode before uploading a finished cue.

Calculate the region from the music, not a rough guess

Our example is an eight-bar instrumental in 4/4 at a constant 120 BPM, with the first downbeat exactly at time zero. Bars five and six contain an overly busy fill. We want a restrained continuation there while keeping bars one through four and seven through eight.

At 120 BPM, one quarter-note beat lasts 60 ÷ 120 = 0.5 seconds. Four beats make a two-second bar. Therefore the eight-bar cue lasts sixteen seconds, and bars five through six occupy 8.0 to 12.0 seconds. The region ends where bar seven begins.

The formula changes when the assumptions change. Add the actual downbeat offset if the file contains leading silence or a pickup. Use a tempo map for music that speeds up or slows down. In meters where the BPM unit differs from a quarter note, use the correct beat unit instead of blindly multiplying by the top number of the time signature.

Distinguish a time selection from an instrument selection

If you supply a stereo mix and mask seconds eight through twelve, you are selecting that portion of the mix. The operation does not inherently mean “replace only the drums while every other instrument stays exact.” Time and instrument are different selection dimensions.

For a drum-only repair, use an isolated drum stem in a compatible workflow and keep the other approved stems in your editor. If you first separate a mixed recording into stems, inspect the separation for bleed and artifacts. The isolation step can introduce its own problems before inpainting begins.

In our example, decide whether the busy fill is the only issue. If the whole arrangement becomes crowded in those two bars, replacing the mixed region may fit the brief. If the guitar melody there must remain exact, preserve it separately rather than relying on an instruction to keep it unchanged.

Describe the replacement’s role in the surrounding phrase

This is a proposed repair brief. It describes what should happen locally and how that passage relates to its neighbors. Avoid a broad genre prompt that invites an entirely different arrangement. You are solving a transition problem, not commissioning a new song inside a four-second hole.

Stability’s current prompt guide covers musical descriptions and model-specific options. Choose a music-capable variant rather than a sound-effects-only one. Tempo and instrumentation instructions remain generation targets; inspect their actual effect rather than assuming the labels guarantee compliance.

Review the two boundaries as musical events

First, play from a little before eight seconds to a little after twelve. Listen for a sudden change in room sound, instrument tone, stereo width, or perceived level. The replacement can be individually pleasant while sounding pasted into another recording.

Next, focus on the incoming beat. Does the new passage land where the previous phrase leads? A clipped cymbal or guitar decay may make the boundary feel abrupt. Then focus on the outgoing beat: the repair must prepare the return to bar seven, not end as if it were a standalone cue.

If the region boundaries cut through meaningful sustained sound, adjust them deliberately. A slightly wider musical selection may provide a better transition, but it also places more approved material at risk. Save the earlier version and compare the wider repair against it.

Audit what was supposed to remain unchanged

Keep the original and repaired files aligned in an audio editor. Compare sections before and after the selected region at matched level. Check duration and start position first; a shifted or resampled export can make a simple comparison misleading.

For a stricter technical check, compare decoded samples outside the region only after confirming identical sample rate, channel layout, alignment, and processing. A null or difference test can help locate changes, but encoding, normalization, fades, and export settings can affect the result. Do not assume every web export preserves outside audio byte for byte.

For ordinary editorial approval, listen to the retained opening and ending as well as the repair. A changed ending is material even if it happened outside the region you intended to select. Keep the source and the final export so that this check can be repeated.

Use a short repair decision table

After accepting a candidate, place it under the actual narration and picture. Fewer notes are useful only if the cue still carries the scene. Review the final compressed delivery file, including its start and tail, before calling the repair finished.

Explore a cue’s initial direction in QuestStudio Music Lab, then use a compatible Stable Audio 3 environment for the inpainting step. This does not imply that Music Lab exposes Stable Audio’s region controls. For cues that must repeat cleanly, the music-loop workflow addresses the separate end-to-start join.

Frequently asked questions