ComfyUI-LavaSR

ComfyUI custom nodes for LavaSR — a fast speech enhancement and audio super-resolution model that upsamples degraded audio to 48 kHz with noise reduction.

Key LavaSR specs:

• 5000× real-time on GPU, ~60× on CPU
• ~50 MB model, ~500 MB VRAM
• Accepts any input sample rate (8–48 kHz)
• Outputs 48 kHz enhanced audio

---

Nodes

LavaSR Model Loader

Loads the LavaSR model from Hugging Face (YatharthS/LavaSR) or a local path. Downloads and caches the model on first use (~50 MB).

Widget	Description
model_name	HF repo ID or local folder path
device	auto picks CUDA → MPS → CPU

Output: LAVASR_MODEL

---

LavaSR Enhance Audio

Enhances a ComfyUI AUDIO tensor. Accepts any sample rate and channel count — audio is downmixed to mono and resampled to 16 kHz internally. Output is 48 kHz mono.

Widget	Description
denoise	Run the denoiser stage before enhancement
batch_mode	Split audio into 1-second chunks (use for long files)
lr_cutoff_hz	Linkwitz-Riley crossover frequency in Hz (default 8000). Set to roughly half your source sample rate

Inputs: LAVASR_MODEL, AUDIO Output: AUDIO at 48 kHz

---

LavaSR Enhance Audio File

Convenience node — loads a file by path and enhances it in one step. Uses the model's native load_audio() which configures the LR crossover automatically.

Widget	Description
audio_file	Absolute path to a .wav file
denoise	Run the denoiser stage
batch_mode	Split into 1-second chunks for long files
input_sr	Sample rate of the source file (default 16000)
cutoff_hz	LR crossover in Hz; 0 = auto (half of input_sr)

Inputs: LAVASR_MODEL Output: AUDIO at 48 kHz

---

Installation

ComfyUI Manager (recommended)

Search for ComfyUI-LavaSR in the Manager and install.

Manual

cd ComfyUI/custom_nodes
git clone https://github.com/marduk191/ComfyUI-LavaSR.git
cd ComfyUI-LavaSR
pip install -r requirements.txt

Restart ComfyUI. The nodes appear under audio/LavaSR.

---

Example Workflows

Three ready-to-use workflows are in the workflows/ folder:

File	Description
basic_enhance_file.json	Simplest — file path → enhance → preview
enhance_loaded_audio.json	Use ComfyUI's LoadAudio picker → enhance → preview
enhance_and_save.json	LoadAudio → enhance → SaveAudio (FLAC)

---

Tips

• TTS output: connect any TTS node's AUDIO output directly to LavaSR Enhance Audio. Leave lr_cutoff_hz at 8000.
• Phone/call audio (8 kHz source): set lr_cutoff_hz to 4000.
• Long files: enable batch_mode to avoid memory spikes.
• Denoise only: set denoise = true and the enhance stage still runs (LavaSR always applies bandwidth extension). To skip it entirely, use the underlying Python API directly.

---

Credits

• LavaSR by Yatharth Sharma — model, training, and architecture
• Vocos — vocoder backbone
• ComfyUI nodes by marduk191