acestep v15 xl sft
简介
核心亮点
- 支持通过自然语言精准生成多样化音效
- 经过 SFT 微调，指令遵循能力强且自然
- MIT 协议开源，企业级集成与二次开发门槛低
- 适用于游戏开发、短视频等音频素材快速产出
使用方法
# 安装 Hugging Face transformers
pip install transformers torch# 使用 transformers 加载模型
from transformers import AutoModel, AutoTokenizer
model = AutoModel.from_pretrained("ACE-Step/acestep-v15-xl-sft")
tokenizer = AutoTokenizer.from_pretrained("ACE-Step/acestep-v15-xl-sft")Hugging Face 下载
我们推荐使用命令行或者 Hugging Face Hub SDK 来进行模型的下载。
操作指引：在下载前，请先通过如下命令安装 huggingface_hub：
pip install -U huggingface_hub命令行下载
下载完整模型库
huggingface-cli download ACE-Step/acestep-v15-xl-sft下载单个文件到指定本地文件夹（以下载 config.json 到当前路径下 ./dir 目录为例）
huggingface-cli download ACE-Step/acestep-v15-xl-sft config.json --local-dir ./dirSDK 下载
# 模型下载
from huggingface_hub import snapshot_download
model_dir = snapshot_download('ACE-Step/acestep-v15-xl-sft')Git 下载
请确保 lfs 已经被正确安装
git lfs install
git clone https://huggingface.co/ACE-Step/acestep-v15-xl-sft如果您希望跳过 lfs 大文件下载，可以使用如下命令
GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/ACE-Step/acestep-v15-xl-sft模型文件托管在 Hugging Face Hub，使用 HF CLI / SDK / Git 直接下载，不经过本站。
PyTorch / Transformers 使用
安装 Transformers
pip install -U transformers torch模型加载和推理
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained('ACE-Step/acestep-v15-xl-sft')
tokenizer = AutoTokenizer.from_pretrained('ACE-Step/acestep-v15-xl-sft')模型下载
我们推荐使用命令行或者 ModelScope SDK 来进行模型的下载。
操作指引：在下载前，请先通过如下命令安装 ModelScope：
pip install modelscope命令行下载
下载完整模型库
modelscope download --model ACE-Step/acestep-v15-xl-sft下载单个文件到指定本地文件夹（以下载 README.md 到当前路径下 dir 目录为例）
modelscope download --model ACE-Step/acestep-v15-xl-sft README.md --local_dir ./dirSDK 下载
# 模型下载
from modelscope import snapshot_download
model_dir = snapshot_download('ACE-Step/acestep-v15-xl-sft')Git 下载
请确保 lfs 已经被正确安装
git lfs install
git clone https://www.modelscope.cn/ACE-Step/acestep-v15-xl-sft.git如果您希望跳过 lfs 大文件下载，可以使用如下命令
GIT_LFS_SKIP_SMUDGE=1 git clone https://www.modelscope.cn/ACE-Step/acestep-v15-xl-sft.gitModelScope 模型页直接下载模型文件；无需将模型文件放在本站服务器。
Notebook 快速开发
下载并安装 ModelScope library
pip install "modelscope[audio,cv,nlp,multi-modal,science]" -f https://modelscope.oss-cn-beijing.aliyuncs.com/releases/repo.html模型加载和推理
from modelscope.pipelines import pipeline
from modelscope.utils.constant import Tasks
p = pipeline('text-generation', 'ACE-Step/acestep-v15-xl-sft')完整文档
---
library_name: transformers
license: mit
pipeline_tag: text-to-audio
tags:
- feature-extraction
- audio
- music
- text2music
- custom_code
---
<h1 align="center">ACE-Step 1.5 XL — SFT (4B DiT)</h1>
<p align="center">
<a href="https://ace-step.github.io/ace-step-v1.5.github.io/">Project</a> |
<a href="https://huggingface.co/collections/ACE-Step/ace-step-15">Hugging Face</a> |
<a href="https://modelscope.cn/collections/ACE-Step/Ace-Step-15-xl">ModelScope</a> |
<a href="https://huggingface.co/spaces/ACE-Step/Ace-Step-v1.5">Space Demo</a> |
<a href="https://discord.gg/PeWDxrkdj7">Discord</a> |
<a href="https://arxiv.org/abs/2602.00744">Tech Report</a>
</p>
Model Details
This is the XL (4B) SFT variant of ACE-Step 1.5 — a supervised fine-tuned model with ~4B parameters. SFT provides higher audio quality with CFG (Classifier-Free Guidance) support for fine-grained prompt adherence control.
XL Architecture
| Parameter | Value |
|-----------|-------|
| DiT Decoder hidden_size | 2560 |
| DiT Decoder layers | 32 |
| DiT Decoder attention heads | 32 |
| Encoder hidden_size | 2048 |
| Encoder layers | 8 |
| Total params | ~4B |
| Weights size (bf16) | ~18.8 GB |
| Inference steps | 50 (with CFG) |
GPU Requirements
| VRAM | Support |
|------|---------|
| ≥12 GB | With CPU offload + INT8 quantization |
| ≥16 GB | With CPU offload |
| ≥20 GB | Without offload |
| ≥24 GB | Full quality (XL + 4B LM) |
All LM models (0.6B / 1.7B / 4B) are fully compatible with XL.
Key Features
- 💰 Commercial-Ready: Trained on legally compliant datasets. Generated music can be used for commercial purposes.
- 📚 Safe Training Data: Licensed music, royalty-free/public domain, and synthetic (MIDI-to-Audio) data.
- 🎯 CFG Support: Fine-tune prompt adherence with guidance scale control.
- 🔮 Highest Quality: SFT + 4B parameters = the highest quality variant.
Quick Start
# Install ACE-Step
git clone https://github.com/ace-step/ACE-Step-1.5.git
cd ACE-Step-1.5
pip install -e .
Download this model
huggingface-cli download ACE-Step/acestep-v15-xl-sft --local-dir ./checkpoints/acestep-v15-xl-sft
Run with Gradio UI
python acestep --config-path acestep-v15-xl-sftModel Zoo
XL (4B) DiT Models
| DiT Model | CFG | Steps | Quality | Diversity | Tasks | Hugging Face | ModelScope |
|-----------|:---:|:-----:|:-------:|:---------:|-------|--------------| ----------- |
| acestep-v15-xl-base | ✅ | 50 | High | High | All (extract, lego, complete) | Link | Link |
| acestep-v15-xl-sft | ✅ | 50 | Very High | Medium | Standard | This repo | Link |
| acestep-v15-xl-turbo | ❌ | 8 | Very High | Medium | Standard | Link | Link |
LM Models (all compatible with XL)
| LM Model | Params | Audio Understanding | Composition | Hugging Face | ModelScope |
|----------|:------:|:-------------------:|:-----------:|--------------| ----------- |
| acestep-5Hz-lm-0.6B | 0.6B | Medium | Medium | Link | Link |
| acestep-5Hz-lm-1.7B | 1.7B | Medium | Medium | Included in main | Included in main |
| acestep-5Hz-lm-4B | 4B | Strong | Strong | Link | Link |
Acknowledgements
This project is co-led by ACE Studio and StepFun.
Citation
@misc{gong2026acestep,
title={ACE-Step 1.5: Pushing the Boundaries of Open-Source Music Generation},
author={Junmin Gong, Yulin Song, Wenxiao Zhao, Sen Wang, Shengyuan Xu, Jing Guo},
howpublished={\url{https://github.com/ace-step/ACE-Step-1.5}},
year={2026},
note={GitHub repository}
}