MD: Audio Auto Master Pro

AI-generated audio comes out of the model raw - fine if you're prototyping, useless if you're shipping. MD_AutoMasterNode is the pack's attempt to do in one node what usually takes a mastering chain: hit a target loudness, fix the spectrum, tame the harsh AI vocal resonances, and leave you with something that doesn't clip or sound like it was generated underwater. Given the pack grew up around ACE-Step music generation, this is the "make it presentable" step at the end of that pipeline.

How it works

It's an iterative mastering engine: analyze the incoming audio with FFT, apply adaptive EQ toward a target spectral profile, run de-essing, multiband compression, stereo widening, and a final brickwall limiter, then normalize to a target LUFS. "Iterative" is the operative word - the max_iterations_eq control (default 5) runs multiple analysis/adjustment passes to converge on the balance.

The one knob that matters most is profile, which sets dozens of DSP parameters at once. Ten profiles ship:

• Standard - balanced all-purpose.
• Diffusion Repair (Clean) - the one the tooltip recommends for raw AI generation output; it's aimed at the specific noise and resonance artifacts ACE-Step leaves behind.
• Podcast (Clarity), Aggressive, Gentle (Tame), Mastering (Transparent), Full Bass (Electronic), Auto-Detect Genre, Custom, and AI Co-Pilot (Ollama).

target_lufs defaults to -14 (streaming standard), -23 for broadcast.

The AI co-pilot

The controversial feature is enable_ai_helper (default True): it queries a local Ollama instance - default http://localhost:11434, default model qwen2.5:14b - and asks it to suggest EQ tweaks based on the analysis, then applies them. A genre_hint text field steers the suggestions. The name is a lie in the best way: it's fully local, no API key, no cloud. The catch is the dependency - if Ollama isn't running, the co-pilot path fails, and you should either start it or turn the helper off and rely on the DSP. For deterministic batch work, disable it; for creative experimentation it's genuinely fun to watch.

Inputs that matter, and what comes out

audio (standard AUDIO dict) goes in. Below the big knobs are surgical controls if you want them: spectral_tilt (macro EQ, ±0.05 is already noticeable - it's extremely sensitive), vocal_tamer_strength (cuts harsh 1–3kHz AI-voice resonances), harmonic_exciter_drive (tube-style warmth, use 0.05–0.20 max), fix_sub_mud_db, fix_kick_thump_db, EQ targets, MBC crossovers/thresholds/ratios, limiter ceiling, stereo_width, and a global mix dry/wet.

Outputs: audio (the mastered result, or the Delta/difference if you set output_mode), plus analysis_details and yaml_config strings, and five plot images - waveform_before, waveform_after, spectrum_plot, dynamics_plot, lufs_history_plot. If you like a result, flip export_yaml and paste it back into yaml_config later to reproduce it exactly.

The honest take

It's a lot of node. The default path (Diffusion Repair + target LUFS) genuinely helps raw ACE-Step output, and the plots make it easy to see what happened. But the profile system is where the value lives - hand-tuning all 30+ parameters is a rabbit hole, and the AI helper is a gimmick until it isn't. Start with the profiles, trust the limiter, and only go manual when the preset doesn't fit.

Installing

It's in 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. The audio stack in requirements (librosa, soundfile, pyloudnorm, pedalboard) is precisely what this node needs, so for once the heavy install is justified.