A video arrives with one mixed soundtrack. You want to replace the dialogue, but keep the background music and sound effects. Muting the audio removes all three.
I built ComfyUI-DBS to bring that separation into a ComfyUI workflow. It splits a soundtrack into music, sound effects, and speech, then lets you save the parts or combine them again. The repository includes a workflow that writes a new video using the music and effects, with the extracted dialogue saved separately.
It runs locally on a CPU or NVIDIA CUDA GPU. The two available models total about 23 MB to download.
This is an unofficial integration of Dynamic Band Splitting (DBS), by Bingzhe Zhang, Kai Li, and Xiaolin Hu. The research and trained models are theirs; my contribution is the ComfyUI integration, workflows, and verification.
DBS produces three audio outputs, often called stems:
For dialogue removal, the workflow adds music + SFX and uses that as the video's new soundtrack. For dubbing preparation or transcription, you can save the speech output on its own. You can also use the separate tracks to adjust their relative levels in a later editing step.
These are estimated sources. Some dialogue can remain in the other tracks, and some wanted sounds can be affected. Start with a short excerpt and listen to every output before processing a longer video.
A 12-second excerpt from Tears of Steel and the DBS-TIGER stems produced by this integration on CPU. Source film: Blender Foundation, CC BY 3.0. The excerpt comes from the DBS project page.
The spectrogram shows the outputs, but listening is the useful test. The authors provide audio examples and a listening demo for exploring the underlying model. Those examples are separate from the ComfyUI integration.
You need ComfyUI v0.38.0 or later, with the V3 node API, and Python 3.10 or later. A CUDA GPU is optional.
Run these commands from your ComfyUI root directory, using the Python environment that runs ComfyUI:
git clone https://github.com/hiroki-abe-58/ComfyUI-DBS custom_nodes/ComfyUI-DBS
python -m pip install -r custom_nodes/ComfyUI-DBS/requirements.txt
If you use a portable installation, replace python with that installation's Python executable. Installing dependencies into a different Python environment will not make them available to ComfyUI.
Download the two models:
python custom_nodes/ComfyUI-DBS/tools/download_model.py --models-dir models/dbs
The download script uses pinned model revisions and checks the files' SHA-256 hashes. It creates these folders:
ComfyUI/
models/
dbs/
DBS-TIGER-DnR/
DBS-BSRNN-DnR/
The requirements are einops and safetensors, which ComfyUI usually already includes. After setup, the nodes load local model files; they do not download models during a run.
Restart ComfyUI to load the new nodes.
Open the supplied video_remove_dialogue.json workflow from:
custom_nodes/ComfyUI-DBS/workflows/gui/video_remove_dialogue.json
Choose your video in Load Video. The workflow's clip.mp4 filename is a placeholder for your own file.
The workflow performs these steps:
add, combines the music and SFX.
The supplied workflow uses channels = each channel to process the stereo channels independently, and output_sample_rate = input to return audio at the input sample rate. The model itself works at 44.1 kHz; the node handles resampling.
For an audio-only input, the GUI workflow folder also includes:
separate_soundtrack.json: saves all three stems as FLAC. remove_dialogue.json: saves music + SFX, plus the separate dialogue track.
You can begin with these workflows and connect the audio outputs to your existing graph.
DBS-TIGER-DnR is the default. It has 1.0 million parameters and higher reported separation scores than DBS-BSRNN on the authors' DnR v2 evaluation.
DBS-BSRNN-DnR has 4.7 million parameters, but it was faster in the repository's local measurements. It is the practical starting point for CPU use.
Here are two examples from the recorded verification runs:
These measurements used Windows 11, an already-loaded model, and the pack's runtime in-process. They exclude model and video export. They are useful reference points, not speed promises for another machine.
Stereo takes more work because each channel is processed separately. For the same 10-minute input on the RTX 5090, the recorded stereo times were 74.5 seconds for DBS-TIGER and 53.7 seconds for DBS-BSRNN. Use those figures when considering the stereo video workflow above.
With the set to auto, the pack uses ComfyUI's GPU when it is CUDA and otherwise falls back to CPU. Apple and AMD GPU acceleration are not implemented in this pack.
A model integration can load successfully while changing the model's behavior. I wanted to check the audio output against the official package as well as make the nodes work.
The verification report records:
Those checks establish integration fidelity under the tested conditions. They do not establish perfect separation quality, or byte-identical output for every exported file and configuration. The report documents the export rounding and CUDA differences too.
The released models were trained on DnR v2, which combines English speech, music, and sound effects into film-like mixtures. Songs with sung vocals, other languages, or unusually loud music need their own evaluation. Singing may be assigned to the music track.
For a first test, choose a short clip with clearly audible dialogue, music, and effects. Listen to the original, the extracted speech, and the recombined music + SFX. That will tell you much more about your use case than a spectrogram alone.
Try ComfyUI-DBS and the included workflows on GitHub. If it is useful in your workflow, a star helps other ComfyUI users discover it. Feedback with your model choice, device, and the kind of audio you tried is welcome in the repository's issues.
ComfyUI-DBS is independent of the DBS authors and Comfy Org. The integration code is MIT-licensed; the vendored DBS code and model licensing are documented in the repository's license notes.