echomom-acestep-v15-xl-sft-4b

ACE-Step 1.5 XL-SFT + 4B Composition LM

This repository is a self-contained checkpoint bundle for ACE-Step 1.5 music generation. It pairs the XL supervised fine-tuned diffusion transformer (approximately 4B parameters) with the 4B 5 Hz composition language model, plus the shared VAE and Qwen3 embedding model required by the ACE-Step pipeline.

The files are an unmodified, revision-pinned assembly of official ACE-Step artifacts. This repository contains no additional training, fine-tuning, LoRA, or custom executable code.

Components and provenance

Model details

The upstream XL-SFT model is the highest-capacity supervised fine-tuned ACE-Step 1.5 DiT variant. It supports classifier-free guidance for prompt-adherence control and normally runs for 50 inference steps. The bundled 4B composition LM is the strongest upstream LM size, intended for song planning, metadata and lyric generation, query rewriting, and audio understanding.

Upstream XL architecture:

Approximate upstream GPU guidance:

Actual requirements depend on precision, duration, batch size, attention backend, and offloading settings.

Repository layout

echomom-acestep-v15-xl-sft-4b/
├── vae/
├── Qwen3-Embedding-0.6B/
├── acestep-v15-xl-sft/
└── acestep-5Hz-lm-4B/

Usage

Use this bundle with the official ACE-Step 1.5 implementation. It is not intended to be loaded as one monolithic transformers.AutoModel; the ACE-Step runtime loads the component directories separately.

git clone https://github.com/ace-step/ACE-Step-1.5.git
cd ACE-Step-1.5
pip install -e .

hf download echomom/echomom-acestep-v15-xl-sft-4b \
  --local-dir ./checkpoints

python acestep --config-path acestep-v15-xl-sft

Consult the upstream project for current installation, inference, Gradio, offloading, and quantization options.

Intended use

This checkpoint bundle is intended for ACE-Step-supported music generation and editing workflows, including text-to-music, cover generation, and repainting. The upstream project describes multilingual prompt support and CFG-based prompt-adherence control for the SFT model.

Users are responsible for evaluating generated content, complying with applicable laws and platform policies, and respecting copyright, privacy, publicity, and other third-party rights. Generated audio may reproduce biases or undesirable patterns present in the upstream training data or model behavior.

License

The repository and pinned upstream model repositories are published under the MIT License. Review the upstream model cards and project documentation for their full usage statements and training-data description.

Acknowledgements

ACE-Step 1.5 is co-led by ACE Studio and StepFun. This repository only assembles their published artifacts into a convenient checkpoint layout.

Citation

@misc{gong2026acestep,
  title        = {ACE-Step 1.5: Pushing the Boundaries of Open-Source Music Generation},
  author       = {Junmin Gong and Yulin Song and Wenxiao Zhao and Sen Wang and Shengyuan Xu and Jing Guo},
  howpublished = {\url{https://github.com/ace-step/ACE-Step-1.5}},
  year         = {2026},
  note         = {GitHub repository}
}