ACE-Step-v1-3.5B / README.md
ChuxiJ's picture
Update README.md
e8177d1 verified
---
license: apache-2.0
tags:
- music
- text2music
pipeline_tag: text-to-audio
language:
- en
- zh
- de
- fr
- es
- it
- pt
- pl
- tr
- ru
- cs
- nl
- ar
- ja
- hu
- ko
- hi
library_name: diffusers
---
# ACE-Step: A Step Towards Music Generation Foundation Model
![ACE-Step Framework](https://github.com/ACE-Step/ACE-Step/raw/main/fig/ACE-Step_framework.png)
## Model Description
ACE-Step is a novel open-source foundation model for music generation that overcomes key limitations of existing approaches through a holistic architectural design. It integrates diffusion-based generation with Sana's Deep Compression AutoEncoder (DCAE) and a lightweight linear transformer, achieving state-of-the-art performance in generation speed, musical coherence, and controllability.
**Key Features:**
- 15× faster than LLM-based baselines (20s for 4-minute music on A100)
- Superior musical coherence across melody, harmony, and rhythm
- full-song generation, duration control and accepts natural language descriptions
## Uses
### Direct Use
ACE-Step can be used for:
- Generating original music from text descriptions
- Music remixing and style transfer
- edit song lyrics
### Downstream Use
The model serves as a foundation for:
- Voice cloning applications
- Specialized music generation (rap, jazz, etc.)
- Music production tools
- Creative AI assistants
### Out-of-Scope Use
The model should not be used for:
- Generating copyrighted content without permission
- Creating harmful or offensive content
- Misrepresenting AI-generated music as human-created
## How to Get Started
see: https://github.com/ace-step/ACE-Step
## Hardware Performance
| Device | 27 Steps | 60 Steps |
|---------------|----------|----------|
| NVIDIA A100 | 27.27x | 12.27x |
| RTX 4090 | 34.48x | 15.63x |
| RTX 3090 | 12.76x | 6.48x |
| M2 Max | 2.27x | 1.03x |
*RTF (Real-Time Factor) shown - higher values indicate faster generation*
## Limitations
- Performance varies by language (top 10 languages perform best)
- Longer generations (>5 minutes) may lose structural coherence
- Rare instruments may not render perfectly
- Output Inconsistency: Highly sensitive to random seeds and input duration, leading to varied "gacha-style" results.
- Style-specific Weaknesses: Underperforms on certain genres (e.g. Chinese rap/zh_rap) Limited style adherence and musicality ceiling
- Continuity Artifacts: Unnatural transitions in repainting/extend operations
- Vocal Quality: Coarse vocal synthesis lacking nuance
- Control Granularity: Needs finer-grained musical parameter control
## Ethical Considerations
Users should:
- Verify originality of generated works
- Disclose AI involvement
- Respect cultural elements and copyrights
- Avoid harmful content generation
## Model Details
**Developed by:** ACE Studio and StepFun
**Model type:** Diffusion-based music generation with transformer conditioning
**License:** Apache 2.0
**Resources:**
- [Project Page](https://ace-step.github.io/)
- [Demo Space](https://huggingface.co/spaces/ACE-Step/ACE-Step)
- [GitHub Repository](https://github.com/ACE-Step/ACE-Step)
## Citation
```bibtex
@misc{gong2025acestep,
title={ACE-Step: A Step Towards Music Generation Foundation Model},
author={Junmin Gong, Wenxiao Zhao, Sen Wang, Shengyuan Xu, Jing Guo},
howpublished={\url{https://github.com/ace-step/ACE-Step}},
year={2025},
note={GitHub repository}
}
```
## Acknowledgements
This project is co-led by ACE Studio and StepFun.