MD: Advanced Audio Preview & Save

If you've ever generated a track with ACE-Step and then fumbled between a DAW and a loudness meter to get a file worth posting, this node is the shortcut. Advanced Audio Preview & Save (the pack calls it AAPS) is the terminal output node for audio workflows: it normalizes to a platform loudness standard, exports to MP3/FLAC/OPUS, embeds your workflow JSON as metadata, and shows before/after waveforms - all in one step.

It's the audio equivalent of a polished image Save node, and it's where most MD Nodes audio pipelines are meant to end.

How it works

The flow is: normalize → effect → encode → save. It reads a standard AUDIO dict, applies a normalization preset, optionally runs a soft limiter, converts channels if asked, applies fades, then writes the file. The two presets do most of the thinking:

• norm_preset - Spotify (-14 LUFS), YouTube Loud (-13), Apple Music (-16), Broadcast EBU R128 (-23), Mastering Headroom (-8), Podcast ACX (-18 RMS), Audiobook ACX (-23 RMS), Peak Normalize Only, or Custom. Pick your platform, done.
• format_preset - High Quality MP3 V0, Web Streaming MP3 128k, Archive/Master FLAC, Voice OPUS 64k, or Music Streaming OPUS 128k.

Set both to Custom and the manual controls unlock: normalize_method (Off/Peak/RMS/LUFS), target_lufs_db, target_rms_db, save_format (mp3/flac/opus), mp3_quality, opus_quality.

The inputs that matter

• save_to_disk - writes the file (False still normalizes and previews, but skips I/O).
• filename_prefix - strftime patterns supported, e.g. audio_%Y%m%d → audio_20251228_00001.mp3.
• save_metadata - embeds the workflow JSON into the file. This is the ComfyUI culture part: drag the file back and the whole graph reconstructs. If the payload exceeds 256KB, it auto-writes a .json sidecar instead. custom_notes goes into a ComfyUI_Notes metadata field.
• channel_mode, fade_in_ms/fade_out_ms - stereo/mono and fades.
• use_limiter - a fast soft limiter at -1.0dB ceiling after normalization; keeps math overshoot from clipping. Keep it on.
• force_save - bypasses ComfyUI's caching so every run writes a new file.

Outputs: AUDIO (pass-through of what was saved), plus waveform_before and waveform_after images - a quick visual confirmation that the normalize didn't wreck the dynamics.

The gotchas

Two real ones. First, LUFS normalization shells out to ffmpeg - the source literally calls ffmpeg -y ... - so if ffmpeg isn't on your PATH, normalization fails. Make sure it's installed before you blame the node. Second, the pack migrated from torchaudio to soundfile for I/O, so the soundfile requirement is load-bearing; if audio I/O errors appear, it's missing from your env. And as always with metadata embedding, a file with embedded workflow JSON is heavier than a bare one - that's the trade for reproducibility.

Installing

It's part of MD Nodes:

cd path/to/ComfyUI/custom_nodes
git clone https://github.com/MDMAchine/ComfyUI_MD_Nodes.git
cd ComfyUI_MD_Nodes && pip install -r requirements.txt

Or via ComfyUI Manager (search MD Nodes), then restart.

Bottom line: if you're shipping AI-generated audio, this node turns "generated a track" into "posted a properly loud, metadata-stamped file" in one shot. The preset system means you rarely touch the manual controls, and the metadata embedding is the exact feature ComfyUI culture expects from a final save node.