---
license: cc-by-nc-4.0
library_name: pytorch
pipeline_tag: text-to-audio
base_model: AndreasXi/MeanAudio
tags:
- text-to-music
- music-generation
- flow-matching
- rlhf
- preference-alignment
---

# TTM-HumanPref: G-Clef FluxAudio-S checkpoint (ICME 2026 ATTM Grand Challenge)

Final submitted checkpoint of **Team G-Clef** for the Efficiency track of the [ICME 2026 Academic Text-to-Music Grand Challenge](https://ntu-musicailab.github.io/ICME26-ATTM-Grand-Challenge/).

A ~120M-parameter FluxAudio-S rectified-flow model, trained from scratch on the challenge's CC-licensed instrumental MTG-Jamendo subset, then refined with human-preference reward conditioning, expert iteration, and a short CRPO preference fine-tune.

- **Paper:** [Improving Text-to-Music Generation with Human Preference Rewards](https://arxiv.org/abs/2606.21670)
- **Code:** [github.com/yonghyunk1m/TTM-HumanPref](https://github.com/yonghyunk1m/TTM-HumanPref)
- **Demo:** [TTM-HumanPref Space](https://huggingface.co/spaces/yonghyunk1m/TTM-HumanPref)

## Files

| File | Size | Description |
|---|---|---|
| `gclef_crpoA_fluxaudio_s.pt` | 483 MB | Final CRPO-A checkpoint, used for both submissions |

SHA-256: `0cfb8295fbd695a99c7a2599363a465fdc28a94f39c8d43faf7487461e5e15f2`

This file is identical to the asset of the GitHub release [`v1.0-icme2026`](https://github.com/yonghyunk1m/TTM-HumanPref/releases/tag/v1.0-icme2026).

## Usage

Inference runs through the code repository, on top of the official challenge baseline. See the [GitHub README](https://github.com/yonghyunk1m/TTM-HumanPref) for setup.

```bash
python run.py download-weights          # fetches this checkpoint and verifies its SHA-256
python run.py reproduce --seed 42       # Submission 1
python run.py reproduce --seed 55       # Submission 2
```

Both submissions come from this single checkpoint and differ only in the generation seed. Inference uses joint classifier-free guidance on text and reward, 25 Euler steps, and the prompt prefix `"high quality instrumental music, "`.

## Intended use and limitations

Research and non-commercial use. The model generates instrumental music only (~10 s, 44.1 kHz) and was trained on a limited instrumental subset of MTG-Jamendo, so its stylistic coverage reflects that data.

## License

The model weights are released under [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/) (see `LICENSE`). The code in the GitHub repository is MIT.

## Citation

```bibtex
@misc{kim2026ttmhumanpref,
  title         = {Improving Text-to-Music Generation with Human Preference Rewards},
  author        = {Kim, Yonghyun and Lee, Junwon and Xia, Haiwen and Ma, Yinghao and Donahue, Chris},
  year          = {2026},
  eprint        = {2606.21670},
  archivePrefix = {arXiv},
  primaryClass  = {cs.SD},
  url           = {https://arxiv.org/abs/2606.21670}
}
```