ComfyUI Spectrogram → Audio

A ComfyUI custom node package that reconstructs audio from spectrogram images using phase retrieval (Griffin-Lim algorithm). Supports both linear magnitude spectrograms and Mel spectrograms.

---

Nodes

Linear Spectrogram → Audio (Griffin-Lim)

Category: audio/spectrogram

Converts a linear or log-magnitude STFT spectrogram image directly to audio.

Parameter	Type	Default	Description
spectrogram	IMAGE	—	Input spectrogram (freq on vertical axis, time on horizontal)
sample_rate	INT	22050	Output audio sample rate (Hz)
n_fft	INT	2048	FFT window size
hop_length	INT	512	STFT hop length (samples)
win_length	INT	2048	STFT window length (samples)
n_iter	INT	60	Griffin-Lim iterations — more = better quality, slower
power	FLOAT	1.0	Exponent applied to magnitude before inversion (1.0 = amplitude, 2.0 = power)
scale	ENUM	log_db	Pixel-to-magnitude mapping (see Scale Modes)
db_range	FLOAT	80.0	Dynamic range in dB (used only when scale = log_db)

Output: AUDIO

---

Mel Spectrogram → Audio (Griffin-Lim)

Category: audio/spectrogram

Converts a Mel spectrogram image to audio via two steps:

1. Inverse Mel filterbank — maps Mel bins back to linear STFT bins
2. Griffin-Lim — recovers phase and synthesizes audio

Parameter	Type	Default	Description
mel_spectrogram	IMAGE	—	Input Mel spectrogram image
sample_rate	INT	22050	Output audio sample rate (Hz)
n_fft	INT	2048	FFT window size
hop_length	INT	512	STFT hop length (samples)
win_length	INT	2048	STFT window length (samples)
n_mels	INT	128	Number of Mel filterbank bins
fmin	FLOAT	0.0	Lowest frequency for Mel filterbank (Hz)
fmax	FLOAT	8000.0	Highest frequency for Mel filterbank (Hz)
n_iter	INT	60	Griffin-Lim iterations
power	FLOAT	1.0	Magnitude exponent
scale	ENUM	log_db	Pixel-to-magnitude mapping
db_range	FLOAT	80.0	Dynamic range in dB

Output: AUDIO

---

Save Audio

Category: audio/spectrogram

Saves an AUDIO tensor to a file in ComfyUI's output directory.

Parameter	Type	Default	Description
audio	AUDIO	—	Audio to save
filename	STRING	output_audio	Output filename (without extension)
format	ENUM	wav	File format: wav or flac

---

Audio Info

Category: audio/spectrogram

Displays metadata of an AUDIO tensor (batch size, channels, sample count, sample rate, duration).

Parameter	Type	Description
audio	AUDIO	Audio to inspect

Output: STRING — formatted info text

---

Scale Modes

The scale parameter controls how image pixel values (0–1) are interpreted as magnitude values.

Mode	Formula	When to use
log_db	pixel → dB ∈ [−db_range, 0] → amplitude	Most visualization tools (e.g. librosa's specshow, Audacity exports)
log_natural	pixel → expm1(pixel × log(2))	Spectrograms saved with log1p normalization
linear	pixel = magnitude	Raw magnitude arrays saved directly as images

Tip: If you generated the spectrogram with librosa.amplitude_to_db() or similar, use log_db.

---

Algorithm — Griffin-Lim Phase Retrieval

Only the magnitude of a spectrogram is stored in an image; the phase is lost. Griffin-Lim reconstructs the phase iteratively:

Initialize: random phase θ₀
Loop n_iter times:
    S  = magnitude ⊙ exp(iθ)      ← impose known magnitude
    s  = iSTFT(S)                  ← synthesize waveform
    S' = STFT(s)                   ← re-analyze
    θ  = angle(S')                 ← update phase estimate
Return: iSTFT(magnitude ⊙ exp(iθ))

This implementation uses momentum=0.99 (Fast Griffin-Lim variant), which converges faster than the original algorithm.

Quality tips:

• Increase n_iter (60 → 128+) for cleaner results at the cost of speed.
• Match n_fft, hop_length, win_length, and sample_rate exactly to the parameters used when the spectrogram was originally created.
• For Mel spectrograms, also match n_mels, fmin, and fmax.

---

Installation

# 1. Clone into your ComfyUI custom_nodes directory
cd /path/to/ComfyUI/custom_nodes
git clone https://github.com/yourname/ComfyUI_spectrogram2audio

# 2. Install dependencies
pip install -r ComfyUI_spectrogram2audio/requirements.txt

# 3. Restart ComfyUI

Requirements

librosa >= 0.10.0
soundfile >= 0.12.0
numpy >= 1.24.0

---

Example Workflows

Linear spectrogram → WAV

[Load Image] ──► [Linear Spectrogram → Audio]
                      sample_rate: 22050
                      n_fft:       2048
                      hop_length:  512
                      n_iter:      60
                      scale:       log_db
                      db_range:    80
                 └──► [Save Audio]
                           filename: my_audio
                           format:   wav

Mel spectrogram → WAV

[Load Image] ──► [Mel Spectrogram → Audio]
                      sample_rate: 22050
                      n_fft:       2048
                      hop_length:  256
                      n_mels:      128
                      fmin:        0
                      fmax:        8000
                      n_iter:      80
                      scale:       log_db
                 └──► [Save Audio]

---

Limitations

• Phase retrieval is an approximation. The reconstructed audio will sound similar to the original but is not identical, especially for complex signals.
• Griffin-Lim works best on single-instrument or speech audio. Mixed or highly transient signals may sound blurry.
• The spectrogram image must preserve the correct aspect ratio (frequency bins × time frames). Aggressive resizing will degrade quality.

---

License

MIT