AI Audio Upsampling

Restoration/Variation of audio using chunked diffusion. Heavily based on Reed+Sam AI's SplitDiffusion implementation of Zach Evans Dance Diffusion. Uses dance diffusion to restore audio in variable chunks. Big shoutout to Reed+Sam Media for coming up with the original idea, it works really great and its fun to mess with.

This is basically a CLI wrapper for the Reed+Sam notebook so you can whip it out and use it on a bunch of files fast.

Installation

Install Torch for your system; preferably the CUDA version if your system can handle it. You technically only need base torch + audio:

Windows:
pip3 install torch torchaudio --index-url https://download.pytorch.org/whl/cu118

Linux:
pip3 install torch torchaudio

After this install the requirements:

pip install -r requirements.txt

Usage

python cli.py ->
  -i INPUT, --input INPUT
                        Input folder to upsample/restore
  -o OUTPUT, --output OUTPUT
                        Output folder to render to
  -c CKPT, --ckpt CKPT  Checkpoint Path (Upsampler/DD model)
  -s SAMPLER, --sampler SAMPLER
                        Sampler type ([)fast, hq, normal, adaptive])
  -t STEPS, --steps STEPS
                        Amount of diffusion steps to take
  -n NOISE, --noise NOISE
                        Noise Level (0-1) to add to process diffusion over
  -l LENGTH, --length LENGTH
                        Chunk Length
  -r RATE, --rate RATE  Sample Rate