ryan5453 / demucs

Demucs is an audio source separator created by Facebook Research.

  • Public
  • 347.4K runs
  • GitHub
  • Paper
  • License
Run with an API
Playground API Examples README Versions

Run time and cost

This model costs approximately $0.062 to run on Replicate, or 16 runs per $1, but this varies depending on your inputs. It is also open source and you can run it on your own computer with Docker.

This model runs on Nvidia A40 GPU hardware. Predictions typically complete within 109 seconds. The predict time for this model varies significantly based on the inputs.

Readme

About

Demucs is an audio source separator model created by Facebook Research. The model ingests an audio file, and returns stems (the separated parts such as the vocals, drums, and bass.)

The original author of Demucs left Meta and therefore the original repository is not maintained, so this Cog uses a fork by the original author. You can find the fork here.

Model Versions

Demucs v4 (Hybrid Transformers Demucs)

  • htdemucs - base model
  • htdemucs_ft - fine-tuned version (this has better results but will take 4x as long)
  • htdemucs_6s - adds piano and guitar stems (piano does not work very well)

Demucs v3 (Hybrid Demucs)

  • hdemucs_mmi - base model

Demucs v2

  • mdx_q - quantized of the base model (trained on only MusDB)
  • mdx_extra_q - quantized version of the base model with extra data