mcpField
F
flux-3-audio-dialogue

flux-3-audio-dialogue

github.com/black-forest-labsCreative & media135152 installsUpdated 9w ago
Repository

About flux-3-audio-dialogue

A skill for directing speech and other sound layers in FLUX 3 video prompts. It covers dialogue, voice direction, sound effects, mixing, silence, and decisions about what to reserve for post-production.

Key features

  1. 01Specify dialogue and speaker
  2. 02Set voice and delivery direction
  3. 03Describe causal sound effects
  4. 04Balance foreground and background audio
  5. 05Choose silence or post-production

Use cases

  1. 01Create a clip with spoken dialogue
  2. 02Keep ambience beneath intelligible speech
  3. 03Plan silent source clips for later audio work

Install

Copy the command and run it in a terminal.

$npx skills add https://github.com/black-forest-labs/skills --skill flux-3-audio-dialogue

After installing, type/flux-3-audio-dialoguein Claude Code to confirm it loaded.

Bundled filesmetadata.json

Details

From the project README

FLUX 3 Audio and Dialogue

Name each layer separately: speech, voiceover, ambience, effects, music, or deliberate silence. One blurred description gives up control of all of them; every sound needs a physical source or narrative role.

Speech. Quote the exact line, name the visible speaker (or label the line voiceover/narration so it is not searching for a mouth to belong to), and add no on-screen text, no subtitles when text is unwanted:

A weather presenter on camera in front of a stylized storm map, speaking directly to
the lens: "Storm season is here, and this time, we're ready." Confident delivery,
clean studio lighting. No on-screen text, no subtitles.

Voice anchors: age range, accent when relevant, register, energy, recording distance. Reusing the same direction preserves a kind of voice, not the same performer across generations.

Speakability. Write for the clip's real duration: short sentences, one thought per line, room before and after the payoff; spell unusual names phonetically; shorten the line before speeding the delivery. A line that cannot finish comfortably needs a shorter script or a longer clip.

Effects are causal, not a detached list:

As the cup hits the tile, it cracks with one sharp ceramic snap.

Mix. Say what leads and what stays under it. Keep background voices out when one line matters:

Her line is foreground and fully intelligible. Café chatter and espresso hiss remain
low and diffuse. A restrained piano pulse enters beneath the final words without
masking them.

Silence and post. Set generate_audio: false for a deliberately silent source clip. Reserve for deterministic post: final loudness, EQ, ducking, and fades; guaranteed wording or speaker identity; frame-accurate sync; subtitles and captions; continuity across separately generated clips.

When a take misses, change one dimension at a time: speaker ownership, line length, delivery anchors, competing layers, action-to-effect causality, or generation versus post.

Comments

Markdown supported