Master Lip Sync: Seedance 2.0 + ElevenLabs
If you've been generating talking or singing characters in Seedance 2.0 and wondering why the lipsync feels close but never quite right, the fix isn't a better prompt for the voice, it's separating the video from the voice entirely. Seedance is excellent at generating mouth movement that matches speech rhythm, but its native voice generation is limited. The workaround is to let Seedance handle the performance and let ElevenLabs handle the voice.
Here's the full workflow, split into talking and singing, because they require different setups.
Talking: Clean Dialogue, No Score
The goal here is simple, you want Seedance to generate a character speaking with natural mouth movement, but you don't want any music or sound effects baked into that audio. A clean vocal track is what lets you swap it out later without fighting a mix.
Prompting for clean speech:
Write your dialogue prompt as you normally would - character, action, emotional tone, but explicitly exclude score and sound effects in the prompt itself. Don't just leave it out; state it. Something like "dialogue only, no music, no background sound" as a constraint line works more reliably than omission.
Keep the performance direction focused on delivery: pacing, breath, emphasis. Seedance is reading these cues to drive the mouth shapes, so vague direction gives you vague lipsync.
Treat this like a scratch track. You're not trying to get a voice you love out of Seedance, you're trying to get timing you love. The words, pauses, and emphasis need to be right because that's what the mouth is animating to.
The swap:
Save the Seedance render.
Export the audio track, this is your reference performance, not your final voice.
Upload the voice file into ElevenLabs' voice changer (speech-to-speech) and convert it to the voice you actually want in the final piece.
Because the ElevenLabs output preserves the timing and cadence of the original performance, it drops back onto the Seedance video without any resync work. The mouth was never animated to a "voice" it was animated to speech patterns so as long as the new voice track follows the same rhythm, the lipsync holds.
This is the whole trick: Seedance gives you correct timing, ElevenLabs gives you the voice identity. Neither tool is being asked to do the other's job.
Singing: Stems In, Reference Out
Singing is harder because Seedance isn't generating the vocal performance from scratch the way it does with dialogue, it needs something to lipsync to. That means the workflow starts before you ever open Seedance.
Building the audio reference:
Generate your song in SUNO as usual.
Export and save the stems: you specifically want the isolated vocal stem, not the full mix.
Find the section you actually need on screen and clip it down. Keep it under 15 seconds. This is a hard practical ceiling right now, not a stylistic choice, longer references degrade the lipsync accuracy or get rejected outright.
That clipped vocal stem becomes your audio reference upload in Seedance.
Generating the performance:
Upload your image reference (the character) alongside the audio reference (the clipped vocal).
In your prompt, describe what the character is singing not just "singing," but the actual lyric content and delivery for that clip, so the generation has something to align the mouth shapes to beyond raw waveform matching.
Because you're feeding it an isolated vocal with no instrumentation bleeding into it, Seedance has a cleaner signal to lipsync against, which is why stem isolation matters more here than it does for the talking workflow.
The 15-second ceiling is the main constraint to plan around. If you've got a full verse or chorus you need covered, you're going to be clipping and generating in segments and stitching them in your editor rather than expecting one continuous sung performance.
What's Coming: Seedance 2.5
Seedance 2.5 is expected to raise that ceiling to a minimum of 30 seconds of singing in a single generation. That changes the segmenting math significantly. A full chorus or a short verse-chorus pass becomes feasible in one pass instead of two or three stitched clips. Worth planning future music video shot lists with that in mind, since it may be worth holding off on chopping longer vocal sections until 2.5 is confirmed stable.