Point-and-Play pretrained models

Weights for the Point-and-Play code release (see its README for installation and assets). Audio examples are on the demo page.

File	Size	Contents
encoder.pt	16 MB	Timbre extractor: LoRA adapters (r = 8, α = 16) for the music-CLAP HTSAT-base audio tower and the CGSP head. The music-CLAP checkpoint itself is downloaded by scripts/download_assets.py.
matched_decoder.pt	1.8 GB	Note synthesizer (EMA weights, 150k steps) trained on timbre codes extracted by encoder.pt. This is the Point-and-Play system.
neutral_decoder.pt	1.8 GB	Note synthesizer (EMA weights, 150k steps) trained on clean music-CLAP embeddings. The paper uses it for the oracle, cue-only and separate-then-clone baselines and for the loss ablation.

Verify the downloads with sha256sum -c SHA256SUMS.

Usage

python -m point_and_play.generate --encoder encoder.pt --decoder matched_decoder.pt \
    --mixture song.wav --offset 30 --text guitar --pitches 48 55 60 67 --velocity 100 --out-dir samples/

The files contain only tensors and plain metadata, so torch.load(path, weights_only=True) works. The decoder files store the model configuration and the latent normalization used in training.

Training

Both the encoder and the matched decoder were retrained with the released recipe:

• Encoder: 18k steps, seed 2. This is the final step, which is also its best validation checkpoint.
• Matched decoder: 150k steps on 8 GPUs × 32 notes.