Audio Models

# Use EmptyLatentAudio node
seconds = 47.6       # Audio duration (1.0 - 1000.0)
batch_size = 1       # Number of latents
# Returns latent: [batch, 64, length]
# Length = round((seconds * 44100 / 2048) / 2) * 2

# Use ConditioningStableAudio node
seconds_start = 0.0   # Start time in composition
seconds_total = 47.0  # Total duration
# Applied to both positive and negative conditioning

# Use VAEEncodeAudio node
# Resamples audio to VAE sample rate if needed (default 44.1kHz)
# Output: latent samples

# Use VAEDecodeAudio node
# Standard decoding

# OR use VAEDecodeAudioTiled node for long audio
tile_size = 512
overlap = 64

# Use LoadAudio node
# Supports: audio files and video files
# Extracts waveform and sample rate
# Returns: {"waveform": tensor, "sample_rate": int}

# Use RecordAudio node
# Live microphone input
# Returns same format as LoadAudio

# Use SaveAudio node
filename_prefix = "audio/ComfyUI"
format = "flac"

# Use SaveAudioMP3 node
quality = "V0"  # Options: V0, 128k, 320k

# Use SaveAudioOpus node
quality = "128k"  # Options: 64k, 96k, 128k, 192k, 320k

# Use PreviewAudio node
# Plays audio in the UI

# Use TrimAudioDuration node
start_index = 0.0     # Start time (seconds, can be negative)
duration = 60.0       # Duration to keep
# Negative start_index counts from end

# Use SplitAudioChannels node
# Input: stereo audio
# Output: left channel, right channel (both mono)

# Use JoinAudioChannels node  
# Input: audio_left (mono), audio_right (mono)
# Output: stereo audio
# Auto-resamples if different sample rates
# Auto-trims to shorter length

# Use AudioConcat node
audio1 = first_audio
audio2 = second_audio
direction = "after"  # or "before"
# Auto-converts mono to stereo
# Auto-resamples to match sample rates

# Use AudioMerge node
merge_method = "add"  # Options: add, mean, subtract, multiply
# Overlays two audio tracks
# Auto-normalizes to prevent clipping

# Use AudioAdjustVolume node
volume = 6   # Decibels (+6 = 2x, -6 = 0.5x, 0 = no change)
# Gain calculation: 10^(volume/20)

# Use AudioEqualizer3Band node
low_gain_dB = 0.0      # Bass boost/cut
low_freq = 100         # Low shelf cutoff
mid_gain_dB = 0.0      # Mid boost/cut
mid_freq = 1000        # Mid center frequency
mid_q = 0.707          # Mid bandwidth
high_gain_dB = 0.0     # Treble boost/cut  
high_freq = 5000       # High shelf cutoff

# Use EmptyAudio node
duration = 60.0        # Seconds
sample_rate = 44100
channels = 2           # 1 = mono, 2 = stereo
# Returns silent audio tensor

ComfyUI/
├── models/
│   ├── checkpoints/
│   │   └── stable_audio/
│   │       └── stable_audio_open_1.0.safetensors
│   ├── vae/
│   │   └── stable_audio_vae.safetensors
│   └── text_encoders/
│       └── sat5_base.safetensors

# Use TextEncodeAceStepAudio node
tags = "electronic, upbeat, synthwave"  # Music tags
lyrics = "..."                          # Song lyrics
lyrics_strength = 1.0                   # How strongly to follow lyrics

# Use EmptyAceStepLatentAudio node
seconds = 120.0
batch_size = 1
# Returns: {"samples": latent, "type": "audio"}

# Use TextEncodeAceStepAudio1.5 node
tags = "jazz, piano, relaxing"
lyrics = "..."
seed = 0                         # For reproducibility
bpm = 120                        # Beats per minute (10-300)
duration = 120.0                 # Song duration
timesignature = "4"              # Options: 2, 3, 4, 6
language = "en"                  # Language code
keyscale = "C major"             # Musical key
generate_audio_codes = True      # Enable LLM (slower but higher quality)
cfg_scale = 2.0                  # Guidance scale
temperature = 0.85               # Sampling temperature
top_p = 0.9                      # Nucleus sampling
top_k = 0                        # Top-k sampling (0 = disabled)
min_p = 0.000                    # Minimum probability

# Use EmptyAceStep1.5LatentAudio node
seconds = 120.0
batch_size = 1

# Use ReferenceTimbreAudio node
conditioning = positive_conditioning
latent = reference_audio_latent  # Optional
# Sets timbre reference for generation
# Turn off generate_audio_codes when using references

ComfyUI/
├── models/
│   ├── checkpoints/
│   │   └── ace_step/
│   │       ├── ace_step_1.0.safetensors
│   │       └── ace_step_1.5.safetensors
│   ├── vae/
│   │   └── ace_step_vae.safetensors
│   └── text_encoders/
│       └── ace_t5.safetensors

{
    "waveform": torch.Tensor,  # Shape: [batch, channels, samples]
    "sample_rate": int         # Samples per second
}