
ByteDance Seedance 2.5 Video Generation Model and Its Audio
What audio-video joint generation in the ByteDance Seedance 2.5 video generation model means for voice, lip sync, lyrics and subtitles.
ByteDance describes Seedance 2.5 as "a next-generation audio-video joint generation model, built for 30-second storytelling with precise reference control and powerful editing capabilities." Most coverage of the model focuses on the picture. The first two words of that description deserve as much attention, because for anyone working with voice, narration or music, joint audio-video generation changes the order in which a video gets made.
This piece looks at the ByteDance Seedance 2.5 video generation model from the sound side: what joint generation means in practice, what it does for dialogue and lyrics, and what it deliberately leaves out. Implementation details come from OiiOii, an AI animation agent platform that runs the model, as confirmed by the OiiOii team on 14 October 2026.
What "audio-video joint generation" means
In a conventional pipeline, picture and sound are separate stages. Animation is produced, then voice is recorded, then lips are matched, then a mix brings it together. Each stage hands off to the next, and each hand-off is a place where timing slips.
Joint generation collapses those stages. The model produces the shot and its audio together, from the same brief and the same references. On OiiOii, audio is generated by default and can be switched off when you want silent footage. It is generation, not a mixing stage: there is no separate track to adjust afterwards.
The practical effect is that sync is a property of the output rather than a job. Lip sync is automatic, with no restriction by mode or model, and there are no manual mouth shapes, phonemes or lip-sync keyframes to set.
How voice and sound steer a generation
References can be audio. Seedance 2.5 on OiiOii accepts up to 50 reference inputs in total, combining images, video and audio. The ceiling is combined, not per type. A sound reference can set the mood of a shot in the same way an image reference sets a face or a palette. As with image references, a small, consistent set works better than a large, contradictory one, because contradictions force the model to choose and it will not choose the same way twice.
Timing is written, not drawn. There are no beat markers, tempo maps or manual time grids. Instead, timestamped instructions place an action or a line at a precise point within the selected duration. Each generation runs from 4 to 30 seconds, chosen by you.
Subtitles follow the audio. Subtitles are fully supported, including automatic lyric subtitles aligned to a song. The text, font, appearance timing and position are all editable, which matters for lyric videos and for any dialogue that will be watched with the sound off.
What it deliberately leaves out
Knowing the boundaries saves generations. On OiiOii there is:
- no manually added score, voiceover or sound effects;
- no volume or mixing control;
- no audio channel settings;
- no placeholder-audio-then-final-audio workflow.
If your process depends on a recorded voiceover or a conventional mix, that work sits outside the model, and the generated track should not be treated as a temporary guide to be replaced. The model is built for directing in words and references, not for editing a waveform.
Planning a voiced project around it
Three decisions are easier to make at the start than to discover at delivery.
Decide whether the generated audio is the final audio. If yes, the model can carry dialogue, lip sync and subtitles in one pass. If no, plan the mix as a separate production step and consider switching generated audio off for the shots that will be re-voiced.
Write the line into the brief. Describe subject, action, camera, setting and mood, then the line or sound and when it happens. For example:
A street musician tunes a battered guitar under a railway bridge, then looks up and sings the first line as a train passes overhead; camera at waist height, slow push-in, sodium streetlight, quiet confidence.
Match the line to the duration. A single sentence stretched to 30 seconds leaves the shot idling. Three lines compressed into 5 seconds will not read. Choose the duration for the action, not the maximum.
Beyond a single shot
Thirty seconds is the ceiling for one continuous generation. Longer pieces on OiiOii run through the story workflow, which generates multiple storyboard shots and assembles them automatically. Its seven agents include a Sound Director, and voice and music are generated as a stage of that workflow after the shot videos. For music videos, the song can come in as an uploaded file, as a link, or as lyrics the system generates the track from.
The resolution is 1080p with 10-bit colour built in, and the aspect ratio is chosen before generation, so a widescreen and a vertical version are two sets of generations rather than one export.
For the full model reference, including access, credits and the complete list of boundaries, read the ByteDance Seedance 2.5 video generation model reference on OiiOii.
Sources
- Model description: ByteDance Seed, Seedance 2.5.
- Implementation details on OiiOii, confirmed 14 October 2026: oiioii.ai.
Cover image generated for this article.
Author
2026/11/12


