DiffErase checkpoints

Pretrained checkpoints for DiffErase (Audio Pirates: Black-box Audio Watermark Removal via Diffusion Priors).

Download into the root of the GitHub repository; the folder layout matches the code:

git clone https://github.com/DiffErase/DiffErase.git
cd DiffErase
hf download DiffErase/DiffErase --local-dir .

Domain	Training data	DiffErase-mel	DiffErase-latent
speech	LibriSpeech	DiffErase-mel/models/differase-mel-speech/	DiffErase-latent/data/checkpoints/differase-latent-speech.ckpt
music	FMA	DiffErase-mel/models/differase-mel-music/	DiffErase-latent/data/checkpoints/differase-latent-music.ckpt
env	Clotho	DiffErase-mel/models/differase-mel-env/	DiffErase-latent/data/checkpoints/differase-latent-env.ckpt

• DiffErase-mel: unconditional diffusers UNet2DModel + DDPMScheduler on 80-band mel-spectrograms (22.05 kHz); audio is reconstructed with BigVGAN.
• DiffErase-latent: unconditional AudioLDM-style latent diffusion (16 kHz, 64-band mel). Each .ckpt contains the UNet, its EMA weights and the frozen VAE. vae_mel_16k_64bins.ckpt and hifigan_16k_64bins.ckpt/.json are the pretrained VAE and HiFi-GAN vocoder from AudioLDM.

Usage:

python remove_differase-mel.py --domain speech \
  --watermarked_dir path/to/watermarked_audio --output_dir path/to/erased_audio --noise_level 0.1