[ComfyUI Load Video] ──► VIDEO ──► [AudioX Video to Audio] ──► AUDIO ──► [Preview Audio]

[VHS Load Video] ──► IMAGE ──► [AudioX Images to Audio (VHS)] ──► AUDIO ──► [Preview Audio]
                  fps ↗ (from VHS Video Info)

Import it via ComfyUI → Load → select the JSON file, or find it in ComfyUI's template browser.

sample workflow

Nodes

| Node | Input | Output | Description | |------|-------|--------|-------------| | AudioX Model Loader | — | AUDIOX_MODEL | Load a local AudioX model | | AudioX Video to Audio | VIDEO | AUDIO | Generate audio from ComfyUI's Load Video node | | AudioX Images to Audio (VHS) | IMAGE | AUDIO | Generate audio from frame sequences (VideoHelperSuite etc.) |

Supported Models

| Model | Notes | |-------|-------| | AudioX-MAF | Recommended — best quality, uses Synchformer visual encoder | | AudioX-MAF-MMDiT | MMDiT variant (in progress, not tested yet) | | AudioX | Base model, no Synchformer (in progress, not tested yet)|

Tasks

Install from ComfyUI Manager Extensions

Search 'ComfyUI-AudioX' in ComfyUI Manager Extensions.

All dependencies will be installed automatically. If you encounter numpy conflicts after installation, see the troubleshooting section below.

Install manually:

1 — Clone the node

cd ComfyUI/custom_nodes
git clone https://github.com/jinxishe/ComfyUI-AudioX.git

2 — Install dependencies

cd ComfyUI-AudioX
pip install -r requirements.txt

Note: torch, torchvision, and torchaudio are not in requirements.txt because ComfyUI already manages them.

3 — Download models

Create the directory structure under ComfyUI/models/AudioX/:

ComfyUI/models/AudioX/
├── clip-vit-base-patch32/       ← shared CLIP (download once)
│   ├── config.json
│   └── pytorch_model.bin
└── AudioX-MAF/
    ├── config.json
    ├── model.ckpt
    └── synchformer_state_dict.pth

# AudioX-MAF (recommended)
huggingface-cli download HKUSTAudio/AudioX-MAF \
    --local-dir "ComfyUI/models/AudioX/AudioX-MAF"

# Shared CLIP model (avoids repeated downloads)
huggingface-cli download openai/clip-vit-base-patch32 \
    --local-dir "ComfyUI/models/AudioX/clip-vit-base-patch32"

Restart ComfyUI after downloading.

Tasks

| Task | Description | custom_prompt required? | |------|-------------|:---:| | V2A — Video to Audio | Generate sound effects matching the video | No | | V2M — Video to Music | Generate background music matching the video | No | | TV2A — Text + Video to Audio | Guide sound effects with a text prompt | Yes | | TV2M — Text + Video to Music | Guide music generation with a text prompt | Yes |

Parameters

| Parameter | Default | Description | |-----------|---------|-------------| | steps | 250 | Diffusion sampling steps. Higher = better quality, slower | | cfg_scale | 7.0 | Classifier-free guidance scale | | sigma_min | 0.3 | Minimum noise level | | sigma_max | 500 | Maximum noise level | | sampler_type | dpmpp-3m-sde | Sampling algorithm | | seed | -1 | Fixed seed for reproducibility; -1 = random |

Notes

Models are trained on 10-second clips. Videos shorter than 10 s are padded with the last frame; the output audio is trimmed to the actual video duration.
GPU memory: ~16 GB VRAM recommended (tested on RTX 4060 Ti 16 GB).
The AudioX Images to Audio (VHS) node requires ffmpeg on the system PATH to assemble frames into a temporary MP4.
CLIP UNEXPECTED key warnings in the log are harmless — they appear because CLIPVisionModelWithProjection loads only the vision head from a full CLIP checkpoint.
Tested on Python 3.12 and CU128.

Troubleshooting

Dependency Conflicts

If you encounter dependency conflicts after installation:

NumPy version conflict (e.g., dctorch requires numpy<2.0.0 or opencv-python requires numpy>=2.0):

# Upgrade numpy to latest version
pip install "numpy>=2.0.0"

This is safe — most packages work fine with numpy 2.x even if they specify older version constraints.

Protobuf conflict (descript-audiotools requires protobuf<3.20):

# Downgrade protobuf if needed
pip install "protobuf<3.20,>=3.9.2"

These dependency warnings from pip are usually safe to ignore if the nodes are working correctly.

Credits

Original model: HKUSTAudio/AudioX — HKUST Audio Lab
Sampling: k-diffusion — Katherine Crowson
Great thanks to ZeyueT/AudioX.

License

MIT — see LICENSE