README.md

---
library_name: pytorch
license: cc-by-4.0
tags:
  - audio
  - music-information-retrieval
  - synthesizer-inversion
  - discrete-diffusion
  - reinforcement-learning
  - grpo
---

# DDSynth-RL checkpoints

Official checkpoints for **DDSynth-RL: Audio Synthesizer Inversion via Discrete Diffusion with Reinforcement Learning**, accepted at ISMIR 2026.

[[Paper]](https://arxiv.org/abs/2608.03032) · [[Demo]](https://ddsynth-rl.github.io/DDSynthRL-Demo/) · [[Code]](https://github.com/DDSynth-RL/DDSynthRL)

## Files

| File | Model | Selection |
|---|---|---|
| `ar_dexed_best_id.pt` | Autoregressive | Best in-domain Dexed validation checkpoint |
| `ar_dexed_best_ood.pt` | Autoregressive | Best OOD NSynth validation checkpoint |
| `fm_dexed_best_id.pt` | Flow matching | Checkpoint used for the in-domain demo; selected on OOD validation |
| `fm_dexed_best_ood.pt` | Flow matching | Best OOD NSynth validation checkpoint |
| `dd_dexed_best_id.pt` | Discrete diffusion | Best in-domain Dexed validation checkpoint |
| `dd_dexed_best_ood.pt` | Discrete diffusion | Best OOD NSynth validation checkpoint |
| `ddsynth_rl_multi_reward.pt` | DDSynth-RL | GRPO with the multi-metric audio reward |
| `ddsynth_rl_clap_crepe.pt` | DDSynth-RL | GRPO continued with CLAP and CREPE rewards |

The demo uses separate checkpoints for in-domain and OOD evaluation of the supervised baselines. The exact selection metric for every file is recorded in `manifest.json`; both released flow-matching checkpoints were selected on OOD validation. Each GRPO checkpoint is used for both evaluation domains.

These are inference release checkpoints: model weights, model configuration, and
training step are preserved exactly, while optimizer and runtime states are
omitted. `manifest.json` records release and source-checkpoint SHA-256 digests,
file sizes, model classes, training steps, and selection criteria. `SHA256SUMS`
can be used for download verification.

## Download

```bash
python -m pip install -U huggingface_hub
hf download MINNE-WU/DDSynth-RL --local-dir checkpoints/ddsynth-rl
```

## Evaluation

From the DDSynth-RL code repository:

```bash
bash scripts/test.sh checkpoints/ddsynth-rl/ddsynth_rl_multi_reward.pt \
  --in-domain-h5 dataset/dexed/test.h5 \
  --ood-audio-root dataset/nsynth/test/audio
```

Dexed must be installed separately at `synth/Dexed.vst3` for render-based evaluation.

## Data

The processed Dexed dataset is not hosted in this repository. Source presets
and the upstream four-way preset augmentation are available from
[SPINVAE-2](https://github.com/gwendal-lv/spinvae2). DDSynth-RL additionally
renders every preset variation under four sampled MIDI note, velocity, and
duration conditions. See the code repository's `DATA.md` for artifact
statistics and the current reproduction boundary.

## Citation

```bibtex
@inproceedings{wu2026ddsynthrl,
  title     = {DDSynth-RL: Audio Synthesizer Inversion via Discrete Diffusion with Reinforcement Learning},
  author    = {Wu, Tristan and Chin, Daniel and Zhang, Junan and Jiang, Junyan and Jing, Yansen and Xia, Gus},
  booktitle = {Proceedings of the 27th International Society for Music Information Retrieval Conference},
  year      = {2026},
  address   = {Abu Dhabi, UAE}
}
```

## License

The checkpoints are released under
[CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). The DDSynth-RL code
is released separately under Apache-2.0.