AI Lip Sync Dialogue Scenes with Seedance 2.5: Prompts and Settings

Dialogue is where AI short dramas most often break. A character can look perfect and move well, and the scene still fails because the mouth does not match the words, the voice changes between shots, or the line is cut off. This guide covers three ways to get AI lip sync dialogue scenes with Seedance 2.5, when to use each, and the exact prompts and settings.

Everything here is taken from the live model pages for Seedance 2.5 Text-to-Video, Seedance 2.5 Image-to-Video and Seedance 2.5 Talking Avatar.

What Seedance 2.5 offers for dialogue

Three Seedance 2.5 models matter for dialogue, and they work differently:

	Text-to-Video / Image-to-Video	Talking Avatar
What drives the speech	Your prompt (and optional reference audio in text-to-video)	Your audio recording
Audio	Native audio generated with the video (generate_audio, on by default)	Follows your recording; the model page notes the generated audio can differ from the original
Inputs	Prompt, reference images, videos and audio (text-to-video); start and last frame (image-to-video)	One portrait image, one audio file, optional prompt
Length	4 to 30 seconds	Follows the audio, up to 120 seconds processed
Resolution	480p to 4k	480p or 720p
List price	$0.36 per second at 720p	$0.22 per audio second at 480p, $0.44 at 720p, 4-second minimum

Prices are the list prices on each model page when this post was written; the Generate button shows the current price before each run.

The rule of thumb: use text-to-video or image-to-video when the scene matters, and Talking Avatar when the exact words matter. A two-character argument in a shop, with movement and a reaction, is a scene. A forty-second confession straight to camera from a recorded voice actor is a performance of exact words.

Method 1: Write the line into the prompt

Seedance 2.5 generates synchronized audio together with the picture. You do not attach a voice file; you write the line in the prompt, in quotes, attached to the character who says it and to a time range.

The model page recommends a three-layer prompt: a one-sentence summary, a timed shot breakdown in 1-second steps, then consistency notes. For dialogue, the timeline is what gives each line room.

Mara from @Image 1 stands behind the counter of the flower shop from @Image 2 and speaks to someone off screen, night drama, medium close-up, static camera at eye level.
0-1s: She sets down her scissors.
1-5s: She looks up and says calmly, "There is no Fourteen. The street ends at twelve."
5-6s: She holds the look, unblinking.
Warm practical light, rain on the window behind her. Rain ambience only, no music, no subtitles.

Settings: aspect_ratio 9:16, resolution 480p while you test, duration 6, generate_audio true, reference image 1 = Mara’s character sheet, image 2 = the shop.

What makes these prompts work:

• One speaker per shot. Two characters talking over each other in one generation is the hardest case. Cut between them instead: a shot for the line, a shot for the reply.
• Leave air. Give a line about a second more than you would need to say it. A clipped last word is the most common failure.
• Keep the mouth visible. Medium close-up or closer, face toward camera, nothing in front of the mouth.
• Write “no subtitles”. The model page lists it as a negative control; otherwise the model may draw text on screen.
• Say “no music” when you will score later. Music generated under the dialogue is hard to remove in the edit.

Keep the voice consistent with reference audio

Seedance 2.5 Text-to-Video accepts reference_audios (up to 30 seconds in total) and lets you bind them to characters in the prompt. The model page’s own example is “@Image 1 and @Image 2 are Character 1, voiced by @Audio 1”. Record or synthesize a short, clean sample of each character’s voice once, then reuse it in every dialogue shot:

Mara from @Image 1, voiced by @Audio 1, speaks to Jun from @Image 2, voiced by @Audio 2, across the counter of the flower shop from @Image 3. Medium two-shot, slow push-in.
0-3s: Jun says, "Then who keeps signing for them?"
3-6s: Mara picks up the slip and says, "Nobody. Nobody lives there."
Warm shop light, rain ambience, no music, no subtitles.

If you need a voice sample, a text-to-speech model such as Seed Speech TTS 2.0 gives you named voice presets plus a voice_instruction for tone and pace. Keep each sample to a few seconds of clean speech without music.

Method 2: Start from an approved close-up

When a dialogue close-up has to match a face exactly (for example the same framing across five episodes of a recurring confession scene), make the frame first and animate it with Seedance 2.5 Image-to-Video. Its prompt guide says the image sets the first frame and the prompt should describe what happens from there, not re-describe what is visible.

0-1s: She draws a breath and glances down.
1-5s: She looks straight into the lens and says softly, "I signed it. I just don't remember doing it."
5-6s: Her eyes fill; she does not look away. Camera holds.
Quiet room tone, distant rain, no music, no subtitles.

Settings: duration 6, resolution 720p, generate_audio true. There is no aspect ratio setting on image-to-video; the output follows the start image, so the close-up must already be vertical for a 9:16 episode.

Method 3: Talking Avatar for recorded lines

Seedance 2.5 Talking Avatar turns one portrait and one audio recording into a speaking video. Only the image and the audio are required. Use it when:

• a voice actor recorded the line and the words and timing must be theirs,
• the speech is long (it processes up to the first 120 seconds of audio),
• or you are fixing a line in a shot that otherwise works.

Recommended inputs, from the model page’s tips:

• Image: one clearly visible speaker, good light, face toward the camera, mouth unobstructed. Use a portrait that already matches the scene, wardrobe and framing you want.
• Audio: clean speech with little background noise or music.
• Prompt (optional): visible behavior only, since the audio supplies the words. Prefer subtle movement.

Maintain a tense, guarded expression, small natural head movements and natural blinking. Keep the camera fixed. She glances once toward the door near the end.

Settings: resolution 480p for previews, 720p for the final. Billing follows the processed audio length, rounded up to whole seconds, with a 4-second minimum.

Talking Avatar shots are performances, not scenes: the camera and background stay simple. Intercut them with wider shots from text-to-video to keep the episode alive.

Fixing a bad dialogue take

When the picture is right and only the mouth or words are wrong, do not throw the take away:

1. Regenerate only the line. Split the shot at the line and regenerate just that part with the same references and the same scene wording.
2. Re-sync the existing clip. A lip-sync model such as ByteDance LipSync audio-to-video takes your existing video plus a corrected audio file and re-times the mouth.
3. Rebuild the close-up. Export a sharp, front-facing frame of the character from the shot (the “export frame” or “save still” command in your editor) and run it through Talking Avatar with the corrected recording.

Option 2 is usually the cheapest when the shot is otherwise approved. Option 3 is best when the actor’s performance needs to change.

For the broader trade-off between native audio and a separate voice-plus-lip-sync pipeline, see native-audio video vs a TTS and lip-sync pipeline. For a comparison of dedicated lip-sync tools, see best AI lip sync APIs.

What about Wan 3.0 and Kling 4.0?

Both can generate audio with the picture too. Wan 3.0 has an optional generate_audio switch and its reference-to-video model accepts reference audio. Kling 4.0 generates synchronized stereo audio and its prompting guide suggests writing “a line of dialogue” into the prompt; its reference-to-video model takes up to three reference audios of 5 to 30 seconds each. If a dialogue shot is also a big action shot, Kling 4.0 is worth a test. For straight conversation, start with Seedance 2.5.

Checklist for every dialogue shot

• One speaker, one line, one shot.
• The line is in quotes, with a time range that leaves a beat at the end.
• The mouth is visible and lit.
• The voice comes from the same reference audio every time.
• “No subtitles” and, if you will score later, “no music”.
• Draft at 480p; render the approved take at 720p or higher.

Dialogue shots then slot into the rest of the episode as described in how to make an AI short drama, and character consistency across those close-ups is covered in keep the same character across AI video shots.

FAQ

Does Seedance 2.5 generate speech and lip sync together? Seedance 2.5 Text-to-Video and Image-to-Video generate synchronized audio with the video when generate_audio is on (the default). Write the line in quotes in the prompt. Check every take, because the exact wording and timing can vary between generations.

How do I keep a character’s voice the same across episodes? Use the same short reference audio for that character in every shot and bind it in the prompt, for example “Mara from @Image 1, voiced by @Audio 1”.

When should I use Talking Avatar instead? When the words and timing must come from a specific recording, or when the speech is longer than a typical shot. Talking Avatar processes up to 120 seconds of audio per request.

Can two characters talk in the same shot? It can work, but one speaker per shot is far more reliable. Cut between speakers as a film editor would.

What languages can the characters speak? The Seedance 2.5 video model pages do not list a language setting. Test a short line in your target language at 480p before you commit an episode to it.

Write your first conversation as one line per shot, then generate it in the video generator with Seedance 2.5 or on the Seedance 2.5 model page.