Home › Blog › seedance not following reference audio

Seedance Reference Audio Not Working? Why It Ignores Your Track (and 3 Fixes)

By the AI Music Video Guide team · Updated July 31, 2026 · 7 min read

This post contains affiliate links — if you sign up through them we may earn a commission at no extra cost to you (disclosure). Researched and edited for accuracy with AI assistance.

Quick answer: Seedance 2.0 treats uploaded audio as a style and rhythm reference, not as a soundtrack to preserve — by default it composes new music that only resembles yours, and it can even re-sing your vocals with different lyrics. Three fixes, in order: (1) reference the file explicitly in your prompt with the @Audio1 tag and assign it a job ("@Audio1 is the finished soundtrack — follow it exactly, do not generate new music"); (2) delete any music description from your prompt text so it can't fight the reference; (3) if it still rewrites your track, use the community black-video workaround — convert the audio section to an MP4 with a black frame and upload it as a video reference, which Seedance follows far more faithfully. And if your actual goal is a character singing your exact vocal, that's a lip-sync job, not an audio-reference job — see our Seedance lip-sync guide.

Why does Seedance 2.0 generate its own music instead of using my track?

Because that's closer to how the feature was built than most people expect. Seedance 2.0 generates audio and video together in a single pass rather than laying your file under the picture (source). An uploaded audio reference influences that generation — rhythm, mood, pacing — but the model composes the actual sound fresh. One audio-reference guide puts it bluntly: using audio directly as a reference "may cause significant differences between the generated audio and the original audio," and with music specifically, "the output may sound similar, but the rhythm and intervals can be different" (source).

So if you uploaded your finished song expecting Seedance to score the video with it, and instead got a soundalike cover in the wrong key — that's not a bug in your account. It's the default behavior, and community testing backs it up: one detailed music-video workflow from May 2026 found that Seedance "didn't lip-sync my reference image, and sometimes it would change the lyrics" when fed audio directly (source). The good news: the behavior is steerable. The fixes below are ordered from cheapest to most involved.

Fix 1: Did you actually tell it what the audio is for? (@Audio1)

Seedance 2.0's reference system is mention-based. You can attach up to 12 files — up to 9 images, 3 videos, and 3 audio clips — and you refer to them in the prompt as @Image1, @Video1, @Audio1 and so on. The critical detail: references that are never mentioned in the prompt, or mentioned without a clear role, tend to get ignored or misread. The guidance from prompt-reference documentation is explicit — "Use @Audio1 for rhythm, voice, or atmosphere" — and references without an assigned role are skipped (source).

So the single most common cause of "Seedance not following reference audio" is a prompt that never mentions the audio at all. Fix it with one sentence:

• Weak: "Cinematic music video of a singer on a rooftop at dusk." (Audio uploaded, never referenced — the model invents its own soundtrack.)
• Strong: "Cinematic music video of a singer on a rooftop at dusk. @Audio1 is the finished song — use it as the only soundtrack and match all cuts to its rhythm. Do not generate new music."

Second half of this fix: check what else your prompt says about sound. If your text asks for an "epic orchestral score" while @Audio1 is a synthpop track, you've given the model two conflicting music briefs, and the text often wins. Strip every adjective about music, genre, or instrumentation from the prompt and let the reference own that job entirely.

What else silently breaks reference audio?

The length rows come from tested guidance: the technical ceiling is 15 seconds but sync quality degrades past 10, so trimming references to 3–8 seconds is the working recommendation (source). The format row deserves more nuance. Host spec pages list MP3 and WAV as officially accepted audio-reference formats, at roughly 15 MB per file and 15 seconds of combined duration (source), while one tested guide reports MP3 as the only format that behaved consistently in its runs, with WAV, AAC, and FLAC sometimes uploading without error and then being ignored (source). We can't confirm the silent-failure pattern independently, so treat re-exporting as MP3 as a cheap first test, not a guaranteed diagnosis. On the upload-slot point we'll stay general, because the UI differs between the official app and the many third-party hosts running Seedance 2.0: before you burn a credit, verify the file is actually listed as an audio reference (some third-party front-ends don't expose an audio upload at all), and confirm your generation length matches the audio you trimmed.

Fix 2: The black-video trick — the community workaround that holds up

When the prompt fix isn't enough — usually when you need the exact track, not a rhythm-matched imitation — the workaround the community converged on is converting your audio into a video file with a black frame, then uploading it as a video reference. Seedance treats the audio track inside a reference video much more literally than a bare audio file: "A black screen video reference makes the output much more consistent with the input audio" (source). The workflow post mentioned earlier reached the same conclusion independently after direct audio references kept changing the lyrics (source).

1. Cut your song into segments. Keep each one at 13 seconds or less — Seedance allows 15, but community testing found longer clips cause sync failures — and use whole-second durations, not fractions (source).
2. Convert each segment to a black-frame MP4. Any editor works (drop the audio on a black canvas in CapCut and export), or one ffmpeg line: ffmpeg -f lavfi -i color=c=black:s=1280x720:r=24 -i clip1.mp3 -shortest -c:v libx264 -c:a aac clip1.mp4 (source).
3. Upload the MP4 as a video reference and your character image as an image reference, then tell the prompt exactly what each is: "@Video1 contains the song — use its audio as the soundtrack. @Image1 is the performer."
4. If your host exposes an audio or sound toggle, make sure it's switched on, and set the generation length to match your segment before generating — the exact control varies between the official app and third-party hosts.
5. Chain the finished segments with Video Extension. Since every generation caps at 15 seconds (as of July 2026), a full song is always a multi-segment build — our guide to making AI videos longer than 10 seconds covers stitching without visible seams.

When the real answer is a lip-sync pass, not a reference fix

Be honest with yourself about what you're asking for. If you want your track under the visuals, the fixes above get you there. If you want an on-screen singer mouthing your exact vocal, no amount of reference-prompting makes that reliable — singing sync is its own problem, sensitive to background instrumentation confusing phoneme detection and to fast or slurred delivery (source). That's a dedicated workflow with its own tricks, and we've written it up separately: how to lip-sync a song in Seedance. If your track came out of Suno, start with our Suno-to-music-video walkthrough instead, since it covers the audio handoff from the start.

Is this worth the credits?

Straight verdict: yes, but budget for retries. Even with the black-video method, expect two to three generations per segment before rhythm and framing both land — community testing suggests capping re-rolls at two per clip before changing your inputs instead of your luck (source). For a three-minute song at 10–13 seconds per segment, that's real credit spend — run your numbers in our credit calculator before committing, and see the full cost picture in what an AI music video actually costs.

One thing the free tier cannot do for you here: at one watermarked, non-commercial credit per 24 hours (as of July 2026), you can sanity-check a single segment a day and nothing more. If you're troubleshooting a whole song, that's not a workable loop — the Basic plan at $9.90/month (as of July 2026) is the realistic floor for iterating, and you can cancel after the project if one video is all you need. If Seedance's 15-second segmenting itself is the dealbreaker for your workflow, it's also fair to look at alternatives with longer native extension — our Kling vs Seedance comparison covers that tradeoff honestly.

Estimate your render cost with our free credit calculator.

Frequently asked questions