If your Seedance clip has the right action but an unwanted score, muting the video also removes its footsteps, atmosphere, and dialogue. Audio separation offers another route: identify the music, inspect what remains, and rebuild the final mix around the usable sound.
On October 13, 2026, filmmaker TomLikesRobots shared this approach for Seedance 2.5 using Meta's SAM Audio. The creator described removing unprompted music while retaining sound effects for a replacement soundtrack. That is a creator report, not a measured success rate. Seeora has not reproduced the example. The workflow below combines documented separation behavior with an original editorial review plan; it does not announce an audio-removal tool on Seeora.
AI-generated concept illustration made with Codex imagegen. It is not a SAM Audio interface, a Seedance output, or a before-and-after test.
What SAM Audio contributes
Meta introduced SAM Audio on December 16, 2025. It separates sounds using text descriptions, visual cues, or time spans. This is an existing audio model applied to a newly observed Seedance workflow, rather than a new Seedance feature announcement.
Meta describes applications including background-noise removal and audio cleanup. It also identifies limitations: highly similar sounds can be difficult to separate, audio itself is not a supported prompt, and complete separation without prompting is outside the model's scope. A busy generated soundtrack should therefore be treated as a candidate for inspection, with no promise that every effect survives intact.
Understand target and residual before exporting
The official SAM Audio repository exposes two outputs: target for the isolated sound, and residual for everything else. If you describe the unwanted music as the target, the residual is the candidate soundtrack to keep. Exporting only the target would retain the music you intended to remove.
For text prompts, the documentation recommends concise lowercase noun or verb phrases. An original starting example is background instrumental music. Use a description of the sound you actually hear; a clip with singing may need a different target description. These are suggested diagnostic inputs, not tested presets.
Visual prompts associate sound with visible objects. Span prompts identify intervals where a target occurs. Background music may have no visible source, so begin with text and use a time span only when your chosen interface supports it and the relevant passage is identifiable.
Start with a representative problem section
Keep the original video and its soundtrack together as an untouched reference. Make a short working copy containing both the unwanted music and something important to preserve: a spoken phrase, footsteps, a door closing, or rain. Include the moment where those sounds overlap, rather than testing only a quiet introduction.
Write down a simple acceptance brief before processing. For a rainy street shot, it might be: remove the orchestral underscore, keep the shoe splash and continuous rain, and preserve the timing of the door slam. For a dialogue scene, list the exact words and the speaker's audible breaths. These notes make a clean-sounding but incomplete result easier to catch.
Use the official demo linked by the creator or a documented implementation you can access. This article does not verify the current demo's login requirements, upload limits, pricing, or export controls. Check those in the service before uploading a full project. Developers can consult Meta's repository for local setup; its example requires checkpoint access and recommends a CUDA-compatible GPU.
Audition both outputs against the original
Listen to the music target first. If it contains a clearly audible word or an important impact, the separation may have taken useful material with it. Then listen to the residual on its own, followed by the residual with the picture. Check at a comparable listening level so a quieter track does not merely seem cleaner.
| Passage | Listen for in the residual | Editorial decision |
|---|---|---|
| Speech under music | Complete words, natural endings, stable voice | Keep only if the dialogue remains usable |
| Footsteps or impacts | Clear attack and natural decay | Replace a missing effect deliberately if needed |
| Rain or room tone | Continuous atmosphere across the whole passage | Repair obvious holes before adding a new score |
| Musical peaks | Recognizable melody or rhythmic leakage | Try a clearer target description or revise the edit |
These checks are proposed acceptance criteria, not reported SAM Audio test results. If one isolated section sounds unstable, compare a revised separation against the same original section. Keep notes about which desired sounds disappeared and which unwanted sounds remained. Avoid stacking repeated processing passes without retaining earlier versions for comparison.
Rebuild the mix without losing synchronization
Place the accepted residual at the original clip's start position in your editor. Keep the untouched soundtrack muted on a separate reference track. Watch a visible contact event near the beginning and another near the end: a foot landing, a door touching its frame, or a mouth closing after a word. If the replacement audio starts late or drifts, resolve the alignment before scoring the scene.
Add your chosen music as a separate track after the retained effects and dialogue pass review. Play adjacent shots together. A consistent score cannot hide a rain bed that disappears at every cut, or footsteps that suddenly change character. Make those transitions intentional with the editor's normal mixing controls, then listen to the exported file as well as the timeline.
Decide whether the shot is worth rescuing
For a clip with no useful dialogue or effects, replacing the entire soundtrack may be simpler. If dialogue is damaged or the important action sound is inseparable from the score, consider rebuilding that sound or selecting another take. Separation is valuable when it preserves something worth keeping; it does not recover an original studio recording that was never present in the generated output.
If your problem begins with an uploaded song changing during generation, read Seedance audio references and original-song checks. That is a different decision from removing a score after generation. For visual finishing, see the post-production VFX explanation.
Before processing the remaining clips, approve one difficult section, record the prompt and chosen output, and save the accepted residual alongside the source. You then have a concrete sound reference for the rest of the sequence.
Documentation reviewed October 14, 2026. Creator claims are attributed; no project benchmark or successful reproduction is implied.