Seed Audio workflow adds lip sync to Seedance 2.0 scenes

• Seed Audio plus Seedance 2.0 can now be used as a practical lip-sync scene pipeline: DavidmComfort's test combined reference voices, storyboards, style transfer, and Seedance video.
• The workflow order is audio first, then character sheets, storyboard cuts, a character-free style plate, and final Seedance assembly, according to DavidmComfort's workflow.
• Seed Audio 1.0 is being pitched as full scene audio: hasantoxr's launch thread says one prompt can generate dialogue, sound effects, and background music together.
• The face-reference block shaped the workaround: DavidmComfort's face-block post shows a likeness/private-information error, while DavidmComfort's workflow says the character sheets had to avoid photorealism.
• Audio-driven Seedance is not limited to dialogue: techhalla's guitar-sync test showed a Leonardo workflow syncing animation to guitar audio.

The key move is generating timed audio before video. hasantoxr's Seedance pairing note says BytePlus is pairing Seed Audio with Seedance by feeding scene audio in as the reference track, so the video follows the timing. DavidmComfort's setting reference adds the production trick: generate the cinematic style plate in gpt-image 2 without the characters, then transfer that look into Seedance.

Seed Audio 1.0 is framed as a scene generator rather than a single-line voice tool. The claim in hasantoxr's launch thread is that one prompt can produce dialogue, sound effects, and background music together.

The feature list in hasantoxr's follow-up breaks the useful bits into three creator-facing pieces:

• Up to 120 seconds of full scene audio per generation.
• Reference audio to keep a character voice consistent across a longer story.
• Multiple characters with distinct voices from one prompt.

Creators are already using the same system for full ebooks, audiobooks, and multi-character scenes, according to hasantoxr's creator-use note. hasantoxr's access note says enterprise API access is open, and individual creators can test it through the BytePlus console.

DavidmComfort's version is a compact AI film pipeline. DavidmComfort's workflow lays out the order:

1. Generate the conversation in Seed Audio using reference voices.
2. Generate character sheets, avoiding photorealistic faces because Seedance was blocking them, at least lately on fal.
3. Generate a rough cartoon storyboard to plan where the characters are and create cuts.
4. Generate a setting and style reference image without characters.
5. Assemble the scene in Seedance 2.0.

The result is a 15-second two-way conversation with cuts and lip sync. DavidmComfort's wrap-up claims the setup solves lip-syncing, two-way conversation, cuts, style transfer into video, and consistent characters.

The consistency layer came from separate assets, not one giant prompt. DavidmComfort's character-sheet post shows the different character references, while DavidmComfort's storyboard post shows the rough board used to structure the cuts.

The style layer was isolated from the cast. DavidmComfort's setting reference says the cinematic setting image was generated in gpt-image 2 with image references, and DavidmComfort's workflow calls out the important constraint: no characters in that style plate.

That separation is the sharpest workflow detail here. The character sheets carry identity, the storyboard carries blocking, and the empty setting plate carries the look.

Seedance 2.0 is also being tested as an audio-driven motion system. techhalla's Leonardo test used Seedance 2.0 for guitar sync and posted a 33-second audio-driven workflow.

The same timing idea appears in hasantoxr's Seedance pairing note: generate the scene audio first, then feed it to Seedance as the reference track. Audio sets timing, video follows.

The workaround exists because face references hit policy friction. DavidmComfort's face-block post shows an API error saying the provided images or videos may contain likenesses of real people or other private information, with content_policy_violation and partner_validation_failed in the response.

The rejected input was not a celebrity prompt. In DavidmComfort's reply, he called the source a “pretty vanilla image,” and his follow-up said the idea was to use older, larger “actors” so the characters would not look like anyone.

DavidmComfort had already been chasing a more filmic Seedance look. the same thread lists a chain of specific prompt template, Nano Banana 2 at 2K, downscale to 1920, AuraSR 4K upscale, Seedance 4K, then a neutral grain pass.