ElevenLabs Music v2.5 Inpainting

Overview

Text to Music v2.5 Inpainting by ElevenLabs edits an existing track without rebuilding the whole song. You upload a song, then provide a structured edit plan that mixes unchanged slices with newly generated sections. The model generates new audio that matches the surrounding context, so your edits land in time and feel coherent. You get a single edited track back as an MP3, ready to review and iterate.

What you can build

• Replace just a chorus with new lyrics while keeping the verse unchanged
• Fix a single awkward line by regenerating a short time span
• Change the style of a bridge (instrumentation, energy, vocal approach) without touching the rest
• Extend a song by adding a new intro or outro around kept audio
• Create a seamless loop by keeping a repeated slice and generating a short transition
• Generate a “similar” continuation by conditioning new sections on a short reference moment

Inputs

• An existing song file to edit, uploaded as an MP3. If the file is longer than 10 minutes, only the first 10 minutes are used.
• A JSON “composition plan” that describes the full edited timeline as an ordered list of chunks.
• For chunks you want to keep unchanged, you specify a start and end time range in milliseconds from the beginning of the uploaded song.
• For chunks you want to regenerate, you provide section text that can include a label (like Verse or Chorus), lyrics lines, and short inline directions (like “guitar solo”).
• For regenerated chunks, you set the target duration in milliseconds. Each new chunk must be between 3 seconds and 120 seconds.
• For regenerated chunks, you provide style tags to include and style tags to avoid. Style tags must be written in English.
• For regenerated chunks, you choose how strongly the new audio should follow the surrounding context (low, medium, or high).
• Optional: for a regenerated chunk, you can point to a short reference slice of the uploaded song (up to 30 seconds) and set how strongly the new audio should match its feel.
• An MP3 output quality preset that sets the track’s sample rate and bitrate.

Outputs

• One edited audio track as an MP3 file.
• The MP3 uses the sample rate and bitrate you selected, and it contains both kept segments and regenerated segments in a single continuous timeline.

Recommended settings

• Keep regenerated chunks musically meaningful. 10–30 seconds per regenerated section is a good starting range for chorus or verse edits.
• Put the most important style guidance early. The first regenerated section has outsized influence on the overall sound.
• Use at least 6–7 style tags in the first regenerated section when you need a specific genre and production direction.
• Use “high” context adherence when you want the edit to blend tightly into the neighboring audio.
• Use negative style tags aggressively to block common failure modes like unwanted vocals, distortion, or genre drift.

Limitations

• The edited timeline can contain up to 30 chunks.
• Total edited duration can be up to 10 minutes.
• Each regenerated chunk must be 3–120 seconds long.
• Any reference slice used to guide a regenerated chunk is limited to 30 seconds.
• Very small kept ranges can be fragile. The underlying inpainting workflow expects real musical context, not single-sample edits.
• The section text used for a regenerated chunk has practical formatting limits. Long, unstructured lyrics blocks may be rejected or produce unstable results.
• Low-quality inputs can reduce edit quality. Heavy MP3 compression, clipping, or noisy mixes can cause audible seams or artifacts around edit boundaries.
• Some inpainting workflows require your track to be stored for editing under your ElevenLabs account. Access may depend on your plan and enabled features.

Safety & compliance

• Uploaded audio is inspected for copyright compliance before it can be used for editing.
• Prompts and section text that include copyrighted lyrics, or that reference specific artists or songs, may be blocked and returned as an error.
• Only upload and edit audio you have the rights to use.
• Audio reference and inpainting guide the sound and structure. They don’t copy or remix the original into a derivative “remaster” by design.