Black Video References Help Preserve Voice Delivery in AI-Generated Scenes

In this workflow, the voice is not an afterthought added to a finished shot. The creator first builds a performance in Eleven v4, then uses that audio to guide Seedance 2.5. The aim is to keep a character’s voice, emotion, and delivery consistent while still allowing the video prompt to control what happens on screen.

The demonstration uses a courier pigeon with Carter, a voice described in the interface as “Charismatic and Relatable.” Audio tags direct the performance line by line: the pigeon begins dazed and catching his breath after a crash landing, clears his throat and turns cocky, then grows exasperated, shouts, calms down, whispers, and ends with a nervous laugh. The script includes lines such as “Twelve blocks. Headwind. And a hot dog guy... a HOT DOG GUY tried to swat me!” The tags act as performance directions, not simply labels for the character’s general mood.

The recommended order is to generate the voice first and storyboard the scene around it. In Flows, the creator places a static scene image and the audio on a timeline, so the planned shot and spoken performance can be considered together before video generation. For a project with multiple characters, each should have a selected voice. The tutorial recommends choosing from the library’s “Great voices for Eleven v4” collection because those voices work well with audio tags and expressive delivery.

The first Seedance method supplies three references: the pigeon character, the rooftop set, and the Eleven v4 voice generation. The creator adds the generated audio to the scene in Flows, then connects the character image, set image, and audio to a Seedance 2.5 generation node. In the prompt, each reference is tagged where it applies: the pigeon lands on the rooftop and delivers the line from the voice reference.

The scene is generated separately from the character image. That gives the prompt room to specify the background, rather than letting the character reference dictate the whole shot. The creator points to a practical reason: an earlier background bike kept turning into a handrail. A separate set reference showing the bike clearly helps constrain that part of the scene.

There is also a timing choice. The voiceover is about 18 seconds, but the creator sets the generated video to about 23 seconds to leave roughly four seconds for the pigeon to crash-land before speaking. The first result is described as good, but not exact: the reference says “I’m good” twice, while the generated scene says it three times. Supplying the audio as an audio reference therefore does not guarantee exact wording.

The second method keeps the same voice recording but changes how it is supplied. The creator converts the MP3 into a plain black MP4 with the audio attached, then uses that file as a video reference instead of an audio reference. In the demonstration, this is done by asking Claude to turn the audio into a black video MP4.

The creator’s explanation is that Seedance 2.5 seems better at extracting the intended delivery from a video file than from an audio file. The black video adds no visual content; it carries the original voiceover in a video container. In Flows, the generation node is duplicated, the audio reference is removed, and the MP4 is tagged in the prompt as the source of the line.

The prompt can still describe a distinct action. In this version, the pigeon crash-lands on the rooftop “with an explosion of feathers,” then delivers the line from the video reference. The creator reports that the delivery is more consistent with the original voice generation than in the first method. That comparison concerns delivery; it does not establish that the generated wording is exact.

The third method adds a transcript with timestamps to the black video reference. In Flows, the creator sends the video to an LLM node with the instruction: “Timestamped transcription of this video broken down to centiseconds.” The resulting transcript gives each line a time, and the creator tags that transcript in the Seedance prompt alongside the black video.

The transcript exposes a timing conflict: the source audio begins speaking at the start of the clip, but the pigeon needs time to crash-land first. The creator adjusts the LLM instruction to add three seconds to every timestamp. The first line then begins at three seconds rather than zero, leaving space for the action before the dialogue.

The practical trade-off is between performance, wording, and timing: the black video is the creator’s preferred way to carry the prepared delivery, while the transcript makes it possible to delay the dialogue so the action can happen first. The final example combines those references, allowing the prompt to focus on the pigeon’s movement and crash landing while retaining the intended voice performance and shifted dialogue timing.