$${\color{white}ComfyUI-}{\color{orange}R}{\color{yellow}T}{\color{white}-HeartMuLa}$$ (v 1.09)

A custom node suite for ComfyUI.

Recommended Loader Settings # $${\color{Green}First \ generate \ 20 \ seconds \ audio \ to \ check \ the \ quality}$$

| GPU Tier | Target Use Case | Quantization | Precision (Mula / Codec) | Compile? | Explanation | | :--- | :--- | :--- | :--- | :--- | :--- | | High-End<br>(24GB+ VRAM)<br>RTX 3090 / 4090 | Highest Quality | none | bf16 / fp32 | True | none quantization loads full weights for max fidelity. fp32 codec ensures zero high-frequency hiss. Compile speeds up batch runs. | | Mid-Range<br>(12-16GB VRAM)<br>RTX 3060 / 4070 | Balanced | 4bit | bf16 / bf16 | False | 4bit saves ~50% VRAM, allowing longer song context. bf16 is faster on modern cards (Ampere/Ada). | | Low-End<br>(8GB VRAM)<br>RTX 2060 / 3050 | High Speed | 4bit | fp16 / fp16 | False | fp16 is universally supported and lighter. Quantization is mandatory here to prevent Out-Of-Memory (OOM) errors. |

---

Key Settings Explained

• quantization:
  • none: Best quality but requires ~16GB+ dedicated VRAM.
  • 4bit (Recommended): Virtually indistinguishable audio quality for most users. Reduces model size significantly, allowing for longer generation durations on standard cards.
• codec_precision:
  • fp32: Critical for studio-quality "High Fidelity" output. It prevents metallic artifacts in high frequencies (cymbals, "s" sounds) but is slower.
  • bf16 / auto: Good for speed and lower memory usage.
• compile_model:
  • True: Adds a 1-2 minute delay on the first generation to optimize the graph. Makes subsequent generations 20-30% faster. Ideal for batch production.
  • False: Instant start. Best for testing single prompts.

---

Installation

Follow these steps to get ComfyUI-RT-HeartMuLa up and running.

$${\color{yellow}Step \ 1: \ Clone \ the \ Repository}$$

Open your terminal or command prompt, navigate to your ComfyUI\custom_nodes folder, and run:

git clone https://github.com/monnky/ComfyUI-RT-HeartMuLa

$${\color{yellow}Step \ 2: \ Navigate \ to \ the \ Directory}$$

Change into the project folder:

cd ComfyUI-RT-HeartMuLa

$${\color{yellow}Step \ 3: \ Install \ Dependencies}$$

First Try install only 3 dep's

pip install soundfile
pip install torchtune
pip install torchao

(if doesnt work, install all requirements) Install the required Python packages:

pip install -r requirements.txt

$${\color{yellow}Step \ 4: \ Download \ Model \ Files}$$

Navigate to your ComfyUI/models directory and download the required weights.

[!TIP] Ensure you have the git-lfs-windows.exe installed. Run the Installer: Double-click the downloaded git-lfs-windows.exe file and follow the setup wizard prompts.

# 1. Go to ComfyUI/models folder and open CMD
cd ComfyUI/models

# 2. Initialize LFS (Crucial for downloading large model files)
git lfs install

# 3. Clone the main Generation repo wrapper
git clone https://huggingface.co/HeartMuLa/HeartMuLaGen HeartMuLa

# 4. Enter the directory
cd HeartMuLa

# 5. Clone the NEW Jan 23, 2026 Components: (Clone components into subdirectories)
# New RL-Tuned Model (3B)
git clone https://huggingface.co/HeartMuLa/HeartMuLa-RL-oss-3B-20260123

# New Codec
git clone https://huggingface.co/HeartMuLa/HeartCodec-oss-20260123

---

$${\color{yellow}optional \ only \ download \ if \ you \ want}$$

# Transcriptor
git clone https://huggingface.co/HeartMuLa/HeartTranscriptor-oss

---------------------------------------------
old models 
git clone https://huggingface.co/HeartMuLa/HeartMuLa-oss-3B
git clone https://huggingface.co/HeartMuLa/HeartCodec-oss

---

Model File Structure Once downloaded, your folder should look like this:

---

$${\color{yellow}Model \ Sources}$$

Explore the technical foundations and official repositories for the models used in this project:

• GitHub Repository: HeartMuLa/heartlib
• Technical Paper: ArXiv: 2601.10547
• Official Demo: HeartMuLa Project Page

Hugging Face Model Hub:

• HeartMuLaGen Config Files: HeartMuLaGen
• Main Model (3B): HeartMuLa-oss-3B
• Codec: HeartCodec-oss
• Transcriptor: HeartTranscriptor-oss

---

$${\color{yellow}Credits}$$

Special thanks to the HeartMuLa team for providing the open-source weights and research that make these nodes possible.

• HeartMuLa Organization: Hugging Face Profile

---

Changelog

[1.09] - 2026-01-24

Updated

• Model Weights: Updated core model and codec files to the latest versions (20260124) for improved adherence and stability.

[1.06]

Fixed

• CFG Scale Crash: Fixed a critical "Batch Size Mismatch" error when switching between CFG 1.0 and higher values. Implemented a hard KV-Cache reset to synchronize batch dimensions.
• Decoder Hang: Fixed an issue where the fp32 codec would hang indefinitely when used with quantized models. Added explicit device casting and eval mode synchronization.
• Linting Errors: Resolved formatting issues in the generation logic for better code stability.

Optimized

• Long-Form Audio: Improved audio clarity for 4-minute tracks by implementing Temperature Decay and aggressive VRAM cache cleanup to prevent robotic artifacts.
• Tag Compliance: Updated conditioning nodes to strictly enforce lowercase and comma-separated tag formatting as required by the paper.

[1.01]

Added

• Split codec support (Dual Precision for Model/Decoder)

[1.00]

Initial Release

• First public release