MelBandRoFormer Audio Separation
Load an audio file and separate it into vocal and instrumental stems using the MelBandRoFormer model. Outputs two MP3 files for isolated vocals and background music.
试试一个示例
关于此工作流
This ComfyUI workflow performs high-quality voice isolation using the MelBandRoFormer model. It takes a single mixed audio track and separates it into two stems: isolated vocals and instrumental accompaniment. The pipeline starts with LoadAudio to ingest your source file, then MelBandRoFormerModelLoader loads the pretrained weights (MelBandRoformer_fp16.safetensors) from models/diffusion_models. The core separation happens in MelBandRoFormerSampler, which runs inference and outputs two audio tensors—one for vocals and one for background music—ready for export via paired SaveAudioMP3 nodes.
Technically, MelBandRoFormer applies a transformer-based architecture that operates in band-split, time–frequency space to estimate source components. In practice, that means it learns masks that suppress accompaniment when extracting vocals and vice versa, yielding cleaner stems with fewer musical artifacts. The included MarkdownNote provides inline guidance inside the canvas. Because the workflow is node-driven, you can easily swap SaveAudioMP3 for a WAV saver, change output filenames, or tweak Sampler options like chunking/overlap (if exposed by your node version) to balance speed, memory use, and separation quality.
常见问答
A modern GPU is strongly recommended for reasonable speeds. Some setups may run on CPU, but processing will be significantly slower. If you encounter memory issues on GPU, reduce any chunk size or increase overlap settings in the Sampler (if available) to manage VRAM.
LoadAudio typically accepts common formats like WAV, MP3, and FLAC. The provided workflow saves MP3 stems via SaveAudioMP3. If you need lossless output, swap the saver node for a WAV variant in ComfyUI.
Try increasing overlap or adjusting chunk size in MelBandRoFormerSampler if those controls are exposed by your node version. Also ensure the input isn’t clipped and avoid overly compressed masters when possible. For final polish, apply light noise gating or spectral denoising to the vocal stem.
This workflow is configured for two stems: vocals and instrumental. For multi-stem separation, use a model and workflow designed for 4-stem or 5-stem outputs, or create a variant that chains models targeting specific sources.