Muse Ambient Audio Generator

Ever wanted a clip of coffee-shop chatter, wind, or crowd murmur to layer under a video without having to find, download, and license an audio file? That's exactly what this node is for. MuseAmbientAudioGenerator extracts the ambient-audio pass that the Director nodes run internally and lets you run it on its own - a pure text-to-ambience generation with no video attached.

The neat part is that it was always text-only under the hood. The Director's ambient pass builds a video latent from an all-zero placeholder (never real rendered frames), so it never depended on a video generation running first. This node just makes that explicit. Give it a description, get back an AUDIO clip.

How it works

The mechanism is a mini LTX audio-sampling pipeline: it tokenizes your description (auto-extracting the [SOUNDS] tag if you paste a full prompt), builds a zeroed video latent plus an audio latent with a full noise mask, runs a small euler sampler pass with the audio VAE's decoder, and hands you back a {"waveform", "sample_rate"} dict in the standard AUDIO format. placeholder_width/placeholder_height exist only to size the dummy video latent - the content is discarded, which is why the tooltip tells you to keep them small.

One thing to understand before you get excited: it generates continuous ambience, not event sounds. There's no visual context, so it can't sync a splash to a splash. You get waves, wind, birdsong, crowd murmur. The source code says as much, and it's a fair description.

Inputs that matter

• sounds_text - the whole ballgame. Plain description ("quiet forest, birdsong, distant stream") or a full prompt with a [SOUNDS] tag, which it auto-extracts. Remember LTX tags must be uppercase.
• base_model (optional) - this is the one people trip on. Wire a plain UNETLoader output (no talking-head LoRA) here. The talking-head LoRA suppresses ambient content, so if you leave this unconnected it falls back to model and you get weak or silent room tone.
• duration_seconds (default 5, max 60), steps (default 12), cfg (default 1), seed - the usual suspects. Distilled flow-matching, so CFG 1.

Output is a single audio socket. Wire it into a preview node, a Video Combine that takes audio, or the bg_audio input of the Director nodes.

Installing it

Same pack as the rest of Muse's LTX tooling - muse-ltx-timeline:

cd ComfyUI/custom_nodes
git clone https://github.com/muse-collective-26/muse-ltx-timeline

Restart, then make sure the Python deps are in place:

pip install av torchaudio soundfile

This node is one of the WIP modules the pack loads in a try/except - if a dependency is missing you'll just see it missing from the node list rather than the whole pack failing to load. And it needs the LTX audio stack: the ltx-2.3-22b-distilled_audio_vae.safetensors in models/vae/, the 22B transformer, and the Gemma 3 text encoder.

Gotchas

• The talking-head LoRA silence trap. If you generate silence, this is why. Feed base_model a LoRA-free UNETLoader output.
• The decoded audio is resampled to 44.1kHz. Fine for room tone, not a mastering tool - if you need a clean mix bus, render it and process in your DAW.
• Keep duration_seconds modest; this is a placeholder-latent pass so it's cheap, but there's no reason to generate a minute of ambience if you'll only use ten seconds of it.

Honestly, for the price of free you get a genuinely useful little tool that slots into the bg_audio track of any Muse Director workflow - and it's the only node in this pack that generates sound with no video attached at all.