version: "1.0.4" name: higgsfield-audio description: > Use when the user asks about audio in Higgsfield videos, needs to add dialogue or lip-sync, wants sound effects or ambient sound in generated video, asks about music or BGM in output, or is using any audio-capable model (Kling 3.0, Seedance 1.5 Pro, Seedance 2.0, Veo 3/3.1, Grok Imagine Video). Also use when the user's prompt would benefit from audio direction but they haven't mentioned it. Also use when the user wants standalone audio — a soundtrack, ambience bed, multi-speaker scene audio (Seed Audio 1.0), or text-to-speech voiceover. user-invocable: true metadata: tags: [higgsfield, audio, dialogue, lip-sync, SFX, ambient, sound, BGM, music, voice, seed-audio, scene-audio, TTS] version: 3.5.0 updated: 2026-08-09 parent: higgsfield

Higgsfield Audio Prompting Guide

QUICK FACTS

Routing aids — read the linked sections for the full rules.

• Native-joint audio models: Kling 3.0, Seedance 2.0 / 1.5 Pro, Veo 3/3.1, Grok — all others add audio in post →
• Four layers to consider per prompt: Dialogue / SFX / Ambient / BGM →
• Lip-sync is the most failure-prone feature: 3–8s clips, MCU framing, one speaking face, locked camera, no head-motion tokens; per-language sync-word budgets are FIELD-reported →
• Seedance 2.0 `@Audio1` is a conditioning INPUT — beat sync, the [AUDIO: Xs] script block, and the first-15s extraction trap →
• Multi-clip assembly: one master track · cuts land on musical punctuation, never inside a sung vowel (ECU mouth-match is the one exception) · unified grain + LUT masks batch color drift →
• Cinema Studio 3.0 native joint audio (SCELA): describe audio as a separate section; specific foley beats generic moods →
• Seed Audio 1.0 (seed_audio, standalone) = whole-scene audio in ONE pass — multi-speaker dialogue + music + SFX + ambience mixed →
• Standalone Audio catalog (2026-08-01 snapshot): seed_audio, qwen_audio_tts (NEW — Qwen 3.0 TTS Flash, expressive instructions + cloned voices), text2speech_v2 (5 engines incl. cozy_voice), plus 3 game-pipeline-only tools — distinct from in-video joint audio →

Which Models Support Audio?

"Native joint" means audio and video are generated simultaneously in one pass — not layered on after. This produces natural synchronization without post-production.

Models without native audio: add audio in post with Lipsync Studio or external tools.

The Four Audio Layers

Every audio-capable prompt should consider four layers. You don't need all four in every prompt, but knowing which to include gives the model clear direction.

1. Dialogue — What characters say

Put dialogue in quotes. Be explicit about who speaks, their tone, and language.

She says: "We need to leave. Now."
He whispers: "Not yet."

Best practices:

• Keep dialogue short — 1-2 sentences per character per shot
• Specify emotional tone: "says urgently", "whispers", "shouts across the room"
• For non-English: specify language and dialect → She speaks in Cantonese: "走啦"
• For Seedance 1.5 Pro: supports English, Chinese (incl. Sichuanese, Cantonese,

Taiwanese Mandarin, Shanghainese), Japanese, Korean, Spanish, Indonesian

2. SFX — Specific sound events tied to action

Describe SFX at the point they happen. Tie them to visible actions.

The glass shatters on the floor — sharp crack, then settling tinkle.
Footsteps on wet concrete — splashing, rhythmic.
A door slams shut — heavy metal, echoing.

Best practices:

• One SFX description per action beat
• Use onomatopoeia sparingly — descriptive phrases work better than "BANG" or "CRASH"
• Tie timing to action: "as she sets the cup down" not "cup sound at 4 seconds"

3. Ambient — Background soundscape

Set the acoustic environment. This is the continuous sound bed.

Ambient: quiet café murmur, espresso machine, rain against windows.
Ambient: forest at night — crickets, distant owl, gentle wind through leaves.
Ambient: busy intersection — traffic, horns, construction in the distance.

Best practices:

• 2-3 ambient elements maximum — more gets muddy
• Describe the space acoustics: "reverberant church hall", "tight car interior"
• Contrast silence with sound for impact: "Dead silence. Then — a single footstep."

4. BGM — Background music mood

Don't name songs or artists (content filter). Describe the musical texture.

BGM: slow piano, minor key, melancholic.
BGM: tense orchestral build — low strings, rising.
BGM: lo-fi hip-hop beat, warm vinyl crackle, relaxed.

Best practices:

• Describe instrumentation, tempo, mood — not genre labels alone
• "Tense strings, building" works better than "suspenseful music"
• Specify when music enters/exits: "Piano enters at the midpoint, builds to the end"
• For beat-sync content: "Cuts match the downbeat" or "Movement peaks on the drop"

Audio Prompt Structure

Add audio cues naturally within your prompt or as a dedicated block at the end.

Inline method (preferred for short prompts):

A woman walks into a quiet library. Her heels click on the marble floor — each step
echoing. She whispers to the librarian: "Do you have the Collected Letters?"
Distant page turns. A clock ticks somewhere above.

Dedicated block method (better for complex audio):

[Scene description — visual content, action, camera]
 
Audio:
  Dialogue: She says "We leave at dawn." He replies: "I'll be ready."
  SFX: coffee cup set down, chair scraping back
  Ambient: early morning kitchen — birds outside, kettle just boiled
  BGM: none — silence emphasizes the tension

Lip-Sync Rules

Lip-sync is the most failure-prone audio feature. Follow these rules strictly:

Expressive facial acting *around* the words — forced smiles, leaking fear,

mixed emotions during a spoken line — is driven separately by FACS Action Unit

codes per beat. Let lip-sync shape the phonemes; schedule the brow/eye/cheek

AUs for the performance. See ../higgsfield-facs/SKILL.md § Dialogue &

Monologue Facial Acting.

Do:

• Keep dialogue clips 3–8 seconds (sweet spot for accuracy)
• Use medium close-up or closer framing — model needs to see the mouth clearly
• One speaking face per shot — multiple faces break audio routing
• Lock the camera: locked-off static camera or slow Dolly In only
• Remove all head/face motion tokens: nodding, turning head, looking around

compete with the lip engine and cause desync

Don't:

• Don't combine dialogue with vigorous head movement in the same prompt
• Don't use 15s clips for lip-sync — technical max but accuracy degrades past 8s
• Don't include ambient or music tokens if lip-sync is the priority — they invite

the generative audio engine to override your dialogue

• Don't use non-MP3 audio for Seedance 2.0 (when available) — WAV/AAC/OGG fail silently

Multi-character dialogue workaround:

Multi-person lip-sync matching is an unresolved limitation across all models. The production workaround:

1. Generate each character separately with their own audio segment
2. Composite in CapCut/Premiere using picture-in-picture + linear mask (15% feather)
3. Static image for the listening character; generated video for the speaking character

Per-language dialogue-sync budgets [FIELD — community, seedance-2.0 repo v6.6.0]

Field-observed word budgets for reliable lip-sync in a ~15s in-video Seedance dialogue clip — not official limits, and not the same as how many words the model can voice. The acoustic budget ≠ reliable-sync budget: the model will happily speak more words than it can keep synced to the mouth.

Cross-language sizing unit: "one short sentence ≈ one breath." Write dialogue in breath-sized sentences and count breaths, not seconds.

Voice-reference lip-sync path [FIELD — community, seedance-2.0 repo v6.6.0]

On surfaces that accept a spoken-voice reference, an attached rights-cleared voice recording drives lip-sync directly — the model syncs the mouth to your recording instead of synthesizing a voice first. This is the most reliable field-reported path for non-English dialogue (it sidesteps the weak-language sync budgets above). Rights-sensitive: only use recordings you have clear rights to — cloned or scraped voices are out.

Audio as a Conditioning Input — Seedance 2.0 (@Audio1)

The most under-used Seedance 2.0 capability: an uploaded audio file is a conditioning input, not just an output track. The model spec lists audio as a reference media role alongside image / video, and generate_audio (native sound output) is documented as independent of the audio reference medias — i.e. the uploaded file conditions the generation, and whether the clip also gets generated sound is a separate switch.

This means @Audio1 has two distinct jobs, and you pick one per shot:

Why it works (author's model — empirical, not in the official spec): the

temporal branch that reasons about motion and pacing reads the sound's

structure — beat positions, dynamic contour, timbral texture, song-structure

sections — and maps it to visual rhythm. Treat the mechanism as a working

model; treat the capability (audio reference role) as confirmed.

Beat sync — the audio choreographs the visuals

Upload an MP3 as @Audio1, then map audio characteristics to visual elements. The minimum is three sentences, each handling one thing — rhythm source / which visual responds / how energy maps to the arc:

Use @Audio1 as the rhythmic foundation. Sync camera transitions to the beat
positions. Visual energy builds with the audio crescendo and peaks at the drop.

You can assign different visual elements to different audio characteristics — mixing audio-to-visual the way you'd mix a track:

@Audio1 drives the visual rhythm. Camera cuts land on the downbeats. Subject
movement accelerates into the build, holds at the peak, releases on the drop.
Colour temperature shifts warmer with the crescendo.

Camera ← beat position. Movement ← dynamic contour. Colour ← overall energy arc.

It stacks with other references — character from @Image1, camera style from @Video1, rhythm from @Audio1, processed together:

@Image1 as character reference. Follow @Video1 camera-movement style. @Audio1 as
rhythmic foundation — sync all camera transitions to the beat positions.
Character movement should pulse with the music.

The one constraint: @Video1 camera style and @Audio1 rhythm have to be temporally compatible. A slow continuous dolly pulled from a video reference fighting an EDM track sends the temporal branch conflicting instructions — same failure class as mixing reference images of clashing styles. Pick references that can coexist. (Sibling of ../higgsfield-seedance/SKILL.md § Reference Roles → Load-Bearing Rule: references stay in their lanes.)

Cutting to music — assembling separately-generated clips on one track

[EMPIRICAL — MiniMax H3 skill corpus, re-derived; cross-model editing craft] Beat sync governs what happens inside a clip; these three laws govern the timeline the clips land on:

• One master track. The piece binds to a single continuous music track laid

in post — never per-clip audio stitched end to end. A join in the music is audible before a join in the picture is visible.

• Cuts land on musical punctuation — a breath, a lyric pause, a snare, the

drop. Never hard-cut inside a sung vowel unless the incoming shot is an ECU whose mouth shape continues that vowel: lip continuity is an edit constraint, not only a prompt constraint.

• Mask batch color drift on purpose. Clips generated in separate runs never

match grade exactly. One unified fine-grain pass plus one LUT across the whole timeline, applied as a deliberate finishing step, hides the inter-clip color variance that would otherwise read as a continuity error.

The [AUDIO: Xs] script block — dialogue + SFX + lip-sync from text alone

No microphone, no recording. A timestamped script inside the prompt text generates voices, SFX, and lip-sync. Quoted text → speech with automatic lip-sync; physical descriptions → sound effects. Each marker is a timestamp in the clip:

[AUDIO: 0s] heavy footsteps on concrete, echoing in a corridor
[AUDIO: 2s] door bursting open, impact bang
[AUDIO: 3s] character says "Nobody move"
[AUDIO: 5s] tense silence, distant traffic
[AUDIO: 7s] character says "Put it down. Slowly."
[AUDIO: 9s] object placed on table, soft thud

The model generates the voice first, then maps facial movement to the waveform — so lip-sync quality is mostly set by how precisely you wrote the dialogue. Exact quoted text outperforms paraphrase. It works across languages (write the line in Spanish/Japanese/French → speech with phoneme-level lip-sync in that language).

This obeys the same physical rules as § Lip-Sync Rules above: a strong @Image1 character reference gives a consistent mouth structure to animate, and close-up framing beats wide (a small face has too few pixels to sync). Keep individual dialogue beats inside the 3–8s accuracy window.

It combines with beat sync in one generation — uploaded music as the rhythmic foundation, the script block as foreground dialogue/SFX, cuts synced to the beat:

@Audio1 as background music. Sync camera transitions to the beats.
[AUDIO: 0s] music from @Audio1 begins
[AUDIO: 3s] character says "This changes everything"
[AUDIO: 5s] sharp breath — beat drop hits simultaneously
[AUDIO: 8s] character says "Let's go"

The 15-second extraction problem — pick the window, don't upload the track

The audio reference limit is 15s, and the model takes the first 15s of whatever you upload. Drop in a full 3-minute track and you almost always feed it the intro — low energy, often ambient, no rhythmic drive. Nothing for the temporal branch to map.

The right 15s follow a build → drop arc: rising tension into a peak. That dynamic gradient is what becomes visual energy structure. A segment with uniform energy gives the model beats to detect but no arc — output is rhythmically synced but dramatically flat.

Where the window lives:

• Pre-chorus into chorus
• Instrumental build into the drop (EDM, electronic, hip-hop)
• Verse climax into a bridge
• The last 15s of an intro that breaks into the first hook

Extract exactly that segment before uploading. MP3 at ≥256kbps — lower bitrate degrades beat detection. Don't upload the full track and hope; pick the window, cut it, upload that. (Flipping the workflow — audio in first, visuals built around it — changes the output at a structural level, not subtly.)

Audio by Model — What Works Best Where

Kling 3.0 (V3) / 3.0 Omni (O3)

• Best overall audio-visual integration
• Multi-language dialogue (English, Chinese, Japanese, Korean, Spanish + regional accents: American English, British English, Indian English)
• Multi-character dialogue: 3+ characters with correct speaker attribution and lip-sync per character
• Voice Binding: lock specific voice profiles to specific characters across shots
• O3 adds Voice Extraction from static images: upload audio clip (min 3s) + image to build a voice profile
• O3 adds Performance Cloning: act out a scene on camera → AI re-renders preserving likeness and voice
• Include dialogue, ambient, and SFX naturally in the prompt
• Prompt like a script: action + camera + mood + dialogue cues together

Audio Speaker Attribution Format (V3/O3):

[Speaker: Character Name] "dialogue" in a [warm/confident/excited] [male/female] voice with [accent].
Add [sound: footsteps / rain / door closing] when [action].
Background ambient: [environment description].

Seedance 1.5 Pro

• Best lip-sync accuracy of all models
• Class-leading multilingual support including Chinese dialects
• Most stable emotional tone control
• Use for: professional dialogue scenes, multilingual content

Seedance 2.0

• Upload MP3 audio as @Audio reference (part of Rule of 12)
• MP3 only — WAV/AAC/OGG/FLAC fail silently with no error
• Max 15s per clip, 3 audio files, 10MB each; ≥256kbps for beat-sync (beat detection)
• Measured enforcement bounds [EMPIRICAL — third-party, China Ark lane, 2026-08]:

per-clip duration is enforced at 1.8–15.2s, and the total across all attached clips is also capped at ≤15.2s — three individually-legal 6s clips get rejected (captured 400 errors). Measured on the China Ark lane by a third party; Higgsfield's own proxy enforcement is unverified — if a multi-clip attach fails, this total cap is the first suspect.

• Timestamp anchoring (audio-as-output): "Audio @Audio1 plays exactly as uploaded from 0s to end. Do not modify."

Then remove all ambient/SFX/music tokens to prevent the generative engine from overriding.

• `@Audio1` is also a visual driver — beat sync, the [AUDIO: Xs] script block,

and the first-15s extraction trap are all in § Audio as a Conditioning Input above.

Diegetic-only convention for the prompt body — a

prompt-authoring discipline that sits on top of Seedance 2.0's

audio capability. BGM is a valid audio layer (see § The Four

Audio Layers above) — that's what Seedance can generate. The

diegetic-only convention is what you should write in the

prompt body: only sounds that physically exist in the scene

(footsteps on wet pavement, fabric whip on motion, breath, room

tone, weather, weapon fire, crowd reaction, stage haze) rather

than naming songs, lyrics, or score cues. If music is intended

for the final cut, layer it in post rather than in the prompt

body.

Two reasons the discipline matters even though BGM is

supported: (i) score descriptors ("dramatic strings",

"orchestral swell") underdetermine the generated audio and

routinely produce generic music beds at odds with the scene;

(ii) the timestamp-anchoring + remove-all-music-tokens

pattern in the bullets above already enforces this discipline

when an MP3 audio reference is uploaded — the diegetic-only

convention generalizes that pattern to all Seedance prompts

whether or not an audio reference is attached.

Veo 3 / 3.1

• Strong native audio for English dialogue and environmental sounds
• Dialogue in quotes: "This must be it," he murmured.
• SFX explicitly: tires screeching loudly
• Ambient as environment soundscape descriptions

Grok Imagine Video

• Improved audio as of Video Imagine 1.0 (Feb 2026)
• Include audio intent directly in prompt — same inline style as other models
• Best for: social clips where audio adds polish but isn't the hero

Common Audio Failures and Fixes

When to Skip Audio

Not every prompt needs audio direction. Skip audio cues when:

• Using a model without native audio (Kling 2.6, Wan 2.6, Seedance Pro, Minimax Hailuo 2.3/02)
• The content is purely visual (product beauty shots, abstract motion, landscape)
• Audio will be added entirely in post-production
• The prompt is already at the short-form 200-word cap and visual direction is more important

Negative constraints: For audio-specific artifacts (lip-sync desync, background music

overriding dialogue, SFX drowning dialogue) and their prevention phrases, see

../shared/negative-constraints.md — Temporal/Consistency Artifacts section.

Cinema Studio 3.0 Audio (Business/Team Plan)

Cinema Studio 3.0 introduces native audio-video joint generation — a fundamental shift from models that treat audio as a post-processing step.

Native Audio-Video Joint Generation

Audio is generated simultaneously with video via a unified multimodal architecture. This means:

• Audio and video are temporally aligned by default — no manual sync needed
• Dual-channel stereo output
• Sound design prompts directly influence both audio AND visual generation
• Audio is not "added on" — it's part of the same generation pass

Audio as Prompt Element (SCELA)

Always describe audio as a separate section in your prompts. The generation engine handles three parallel audio tracks:

1. BGM — background music, score
2. Ambient SFX — environmental sounds, foley
3. Dialogue — character speech, voiceover

A chef slices vegetables rapidly on a wooden cutting board.
Camera: tight close-up tracking the knife.
Style: warm kitchen lighting, shallow depth of field.
 
Audio: rhythmic chopping on wood, oil sizzling in a nearby pan,
soft clinking of ceramic bowls. Light acoustic guitar BGM.

Input Constraints

Lip-Sync

Available but experimental in Cinema Studio 3.0:

• Focus on emotion and general mouth movement, not perfect phoneme sync
• Single face per generation only — multi-face lip-sync is not supported
• For best results, keep dialogue segments under 8 seconds
• Pair with clear frontal or 3/4 face angle reference images

Tone / Voice Cloning via @Reference

Control speaking style, accent, and language by referencing a video with the desired voice:

Voiceover tone references @Video1. The narrator describes the product
in a warm, conversational tone. "This changes everything."
Audio @Audio1 plays exactly as uploaded from 0s to end.
Do not modify or replace the audio content.

Dialect Support

Dialects written directly in the prompt work — the model understands regional speech patterns. Write dialogue in the target dialect for authentic delivery.

Timestamp Anchoring

When uploading reference audio that must play unmodified:

Audio @Audio1 plays exactly as uploaded from 0s to end.
Do not modify or replace the audio content.

Then remove all ambient/SFX/music tokens from the prompt to prevent the generation engine from overriding the uploaded audio with generated sound.

Sound Design Specificity

Describe specific foley, not generic moods:

Wrong: nice ambient sounds, pleasant background noise

Right: the scratch of frosted glass, rustling of plush fabric, gentle tapping on acrylic, popping of bubble wrap, wooden floor creaking under bare feet

Specific sound descriptions directly influence the generated audio output. The more precise the foley description, the more accurate the result.

Scene-Audio Generation — Seed Audio 1.0

Separate from in-video joint audio above: Seed Audio 1.0 (ByteDance, released 2026-06-23 at the FORCE conference) is a standalone one-pass whole-scene audio generator. One generation produces multi-speaker dialogue + music + SFX + ambience, already mixed — a radio-drama scene, not a single voice track. Use it to build a soundtrack for footage you'll assemble in post, or scene audio that has no video at all.

When to choose it — decision table

Verified surface [OFFICIAL — model spec 2026-08-01]

Model id seed_audio (output_type audio). Parameters:

Media roles: image_references + audio_references. Per the fal schema: up to 3 reference audio clips (each ≤30s, ≤10MB) XOR one image reference — image and audio refs cannot combine. Reference audio inputs in the prompt by load order: @Audio1, @Audio2, @Audio3.

Script-format prompting [EMPIRICAL — community guides, NOT official docs]

Everything in this subsection is community-converged practice, not spec — treat as a starting point, not a guarantee. Write the prompt as a radio-drama script:

• Open with a scene header: [Scene: busy coffee shop, morning]
• Speaker labels with emotion parentheticals: Host (warm, upbeat): "…"
• Inline sound cues where they happen: [sound: espresso machine, soft jazz fades in]
• Music by mood, not genre: "soft piano builds to triumphant orchestra", not "cinematic score"
• ~4 distinct speakers is the reliable ceiling
• Test 20s segments before scaling toward the ~2-minute cap
• Output WAV (format: wav) when the audio is headed for post

Compact worked example:

[Scene: rain-soaked night market, closing time]
Vendor (tired, warm): "Last skewers — half price, take them."
Girl (excited): "Two! No — three!"
[sound: rain drumming on tarp canopy, a scooter passing in the distance]
Vendor (chuckling): "Three it is. Careful, they're hot."
[sound: coins dropped on a metal tray, charcoal hiss]
Music: a lonely muted trumpet fades in under the rain, wistful but hopeful.

Standalone Audio tab — tool catalog (2026-08-01 snapshot)

The live standalone-audio catalog, reconciled against the models_explore snapshot of 2026-08-01 (../../specs/models_explore_snapshot_audio_2026-08-01.json; generated table: ../../specs/AUDIO-MODEL-SPECS.md, machine twin ../../specs/audio-model-specs.json — regenerate with python3 scripts/sync_specs.py --type audio). The Audio tab's UI tools — Voiceover (text → speech), Change Voice (swap a voice in any video), Translation (translate speech in any video) — sit on top of these models:

Engine picks within text2speech_v2: seed_speech when the deliverable is multilingual voiceover/narration; elevenlabs (Eleven v3) when fine emotional/tone control matters; vibe_voice for long-form narration. These are standalone audio generators — distinct from the native joint audio baked into Kling 3.0 / Seedance 2.0 / Veo during video generation. (Catalog reflects the 2026-08-01 snapshot; verify live before quoting pricing or availability.)

Post-generation voice-over — Supercomputer workflow [DEMO]

A post-generation alternative to prompting audio at all: upload the finished clip to Supercomputer and ask for an analyzed voice-over (e.g. "Analyze the video and create a voiceover for it in the style of wildlife documentaries") — the agent analyzes the footage, writes a script, and offers voices to pick from. Shown working in Higgsfield's Seedance-4K tutorial; useful when the visuals are already locked and only narration is missing.

Related skills

• higgsfield-seedance-vfx — Footage transforms whose payoff is a camera move synced to a spoken line (crash-zoom / push-in), or preserving the source talk track through a transform (SFX and source dialogue only); see ../higgsfield-seedance-vfx/references/dialogue-timing.md
• higgsfield-models — Which models support native audio
• higgsfield-troubleshoot — Audio failure diagnosis
• higgsfield-cinema — Cinema Studio audio workflow with Kling 3.0
• higgsfield-vibe-motion — Motion graphics with audio (different from AI-generated audio)