---
license: apache-2.0
pipeline_tag: text-to-audio
tags:
- text-to-audio
- sound-effects
- foley
- flow-matching
- diffusion-transformer
- dac-vae
- english
- chinese
- moss
- mirror
---

# MOSS SoundEffect v2.0 — inference mirror for MAESTRO

Inference-ready repack of
[OpenMOSS-Team/MOSS-SoundEffect-v2.0](https://huggingface.co/OpenMOSS-Team/MOSS-SoundEffect-v2.0)
(OpenMOSS, Fudan University), kept by AEmotionStudio so the MAESTRO DAW can
fetch the checkpoint on demand. The model is **not** fine-tuned or otherwise
altered — this repo only repacks the upstream weights for loading:

* `transformer/diffusion_pytorch_model.safetensors` — the 1.3B DiT, stored
  in the inference dtype MAESTRO pins after GPU validation (upstream ships
  fp32 and runs under bf16 autocast).
* `vae/diffusion_pytorch_model.safetensors` — the 48 kHz continuous DAC VAE,
  converted from the pickled `vae_128d_48k.pth` (audiotools package) to
  safetensors; the constructor kwargs live in `vae/config.json`.
* `text_encoder/` + `tokenizer/` — the Qwen3 text encoder shards and
  tokenizer, verbatim from upstream.
* `model_index.json`, `scheduler/`, `transformer/config.json` — verbatim.

| Component | Purpose |
|---|---|
| `transformer/` | 1.3B flow-matching Diffusion Transformer |
| `vae/` | Continuous DAC VAE decoder (latents → 48 kHz mono waveform) |
| `text_encoder/` + `tokenizer/` | Qwen3 text encoder (EN + ZH prompts) |
| `scheduler/` | Flow-match scheduler config (shift 5.0) |

MOSS SoundEffect generates **48 kHz sound effects up to 30 s** from English
or Chinese text prompts — nature, urban, creature, human-action, and
musical-element sounds.

## License

* **Code**: Apache-2.0 ([OpenMOSS/MOSS-TTS](https://github.com/OpenMOSS/MOSS-TTS),
  `moss_soundeffect_v2`).
* **Weights (this repo): Apache-2.0** — inherited unchanged from upstream.

Credit for the model belongs entirely to the OpenMOSS team at Fudan
University; this mirror exists only for reliable on-demand delivery into
MAESTRO.