Will3024/ctrl-lfm-guitar-synth

guitar-synth checkpoints

Pretrained weights for ctrl-lfm-guitar-synth (code + listening demo): score (GuitarPro / MusicXML / MIDI) -> expressive performance curves -> audio, for solo acoustic/electric guitar. Three frozen models chained together, plus one optional post-process and one baseline/ablation.

Files

file	stage	trained by	config
checkpoints/gain/note_gain_v5_samestring_fixed.pt	1. note-gain predictor (score -> per-note loudness)	scripts/perf/train_note_gain_predictor.py	config/perf.yaml
checkpoints/perf/pe_note_shape_fixed13.pt	2. envelope regressor (score + loudness -> pitch/envelope curves)	scripts/perf/train_perf_predictor.py	config/perf.yaml
checkpoints/synth/joint_synth_peov_embed.pt	3. joint six-string synth (curves -> audio latents, rectified flow)	scripts/synth/train_joint_synth.py	config/synth.yaml
checkpoints/reverb/mix_reverb.pt	optional post-process (six dry strings -> one mic-domain mix)	scripts/synth/train_mix_reverb.py	--
checkpoints/synth/joint_synth_velocity_peak.pt	baseline/ablation: single-stage synth conditioned on flat MIDI pitch + peak pseudo-velocity only	scripts/synth/train_joint_synth.py	config/synth-velocity-baseline.yaml

These are the exact weights the repository's own training commands reproduce (see each script's # Run as: header and each config's header comment).

Usage

pip install -r requirements.txt   # from the code repo
python -m scripts.download_checkpoints --repo-id Will3024/ctrl-lfm-guitar-synth
CUDA_VISIBLE_DEVICES=0 python -m inference.infer --gp examples/008_PSswc.xml --out runs/demo

Or fetch a single file directly:

from huggingface_hub import hf_hub_download
hf_hub_download(repo_id="Will3024/ctrl-lfm-guitar-synth", filename="checkpoints/synth/joint_synth_peov_embed.pt")

License

CC BY 4.0, matching GuitarSet, the dataset these were trained on. Use, modify, and redistribute, including commercially, provided you give attribution -- cite this repository and GuitarSet:

Q. Xi, R. Bittner, J. Pauwels, X. Ye, and J. P. Bello, "GuitarSet: A Dataset for Guitar Transcription", ISMIR 2018.

Upstream dependency you must separately accept. Running inference with any of these checkpoints requires downloading stabilityai/stable-audio-open-1.0 at runtime (the frozen latent encoder/decoder, SAO3). That model is governed by the Stability AI Community License, not this repository's license: free for organizations/individuals under $1M USD annual revenue, subject to Stability's Acceptable Use Policy; above that threshold requires a separate Stability AI Enterprise License. You must accept Stability's license yourself on Hugging Face (huggingface-cli login + accepting the gated model) before HF_TOKEN can download it. Generated audio's downstream use is bound by Stability's Acceptable Use Policy in addition to this repository's own terms.

Citation

@software{fu_ctrl_lfm_guitar_synth,
  title        = {ctrl-lfm-guitar-synth: Expressive Guitar Synthesis via Controllable Latent Flow Matching},
  author       = {Fu, Hsueh-Wei},
  affiliation  = {National Taiwan University},
  url          = {https://github.com/WeiFu3024/ctrl-lfm-guitar-synth}
}