IAMCCS Cine Audio Prompt Director

Audio-guided video models are picky about prompt structure - jam scene, camera, dialogue and voice direction into one soup and the model plays favorites with whichever part you emphasized. IAMCCS_CineAudioPromptDirector (a Cine-branded subclass of the pack's LTX-2 AudioPromptDirector) is a small assembly line that takes the four parts of an audio scene separately and builds one clean prompt, plus a ready-to-use dialogue_clause.

Inputs, and they're self-explanatory: scene_prompt, action_prompt, dialogue_text, voice_direction, and speaker_label ("the character" by default). Then three knobs shape the output:

• audio_role - off, dialogue_only, dialogue_plus_action (default), or ambience_only. This controls whether the dialogue gets quoted into the prompt at all and whether the audio is treated as speech or as ambience/timing texture.
• prompt_style - cinematic (reads like a director's note), literal_ltx (explicit Scene: / Action: / Audio: sections, which some LTX-2 chains follow more obediently), or compact (just the parts, no framing).
• fps - used to compute frame counts from the optional audio input.

Outputs: prompt (the assembled string for your text encoder), dialogue_clause (just the "Speaker says '…' with voice direction" fragment, in case you want it separately), audio_seconds / audio_frames (measured from the optional AUDIO input, 0 if none), and report.

The mechanism is the interesting part: it builds the clause like {speaker} says "{dialogue}" with {voice}, then assembles the full prompt per style, always appending a line that tells the model the voice follows the audio while the blocking/camera/cuts follow the written direction. That split - "voice from audio, motion from text" - is the core idea, and it's the right way to prompt an audio-guided model. You can also feed an AUDIO clip in and it'll report its duration, which is a nice "is my clip the length I think" freebie.

Installing: part of IAMCCS/IAMCCS-nodes - ComfyUI Manager → search "IAMCCS", or cd ComfyUI/custom_nodes && git clone https://github.com/IAMCCS/IAMCCS-nodes.git, restart. No models, no extra deps.

The trap: audio_role: off still works and just produces a scene/action prompt - but if you do have dialogue and you leave it off, you'll wonder why the model isn't saying the words. And the difference between cinematic and literal_ltx is not cosmetic for finicky chains: if your sampler runs a stock LTX-2 workflow, literal_ltx tends to track better; if you've got a Cine/FLF setup, cinematic reads more naturally. Try both - the node exists to make that a two-click experiment instead of a rewrite.