ACE-Step 1.5 XL SFT

ACE-Step 1.5 XL SFT

Generates music from text with a 4B diffusion decoder, 50 steps and guidance control, aiming for higher audio quality than the base. For creators with a GPU of 12 GB or more.

Made by ACE-Step on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This is the XL SFT variant of ACE-Step 1.5, an open text to music model. SFT means it was further trained on curated examples, and the card says this gives better audio quality than the base. It has about 4B parameters in its diffusion decoder, supports classifier-free guidance and takes 50 inference steps. ACE Studio and StepFun co-lead the project.

Who it suits: Musicians, creators and developers who want to generate music from text on their own GPU, and who care about audio quality more than speed.

Industry | Who uses it | How they use it | Example
Media | Video creators | Generate background music from a text description | A calm track for a vlog intro
Music | Songwriters | Sketch musical ideas before recording | A rough demo in a chosen style
Games | Indie developers | Make music loops for a project | A looping track for a menu screen
Software | Developers | Build a music tool around an open model | An app that turns prompts into tracks

You can skip it if you need quick drafts, since a turbo variant in the same family needs only 8 steps.

Where to use it: It runs in Python on Windows, macOS or Linux with a GPU. The card lists memory needs from 12 GB of VRAM upward.

• The ACE-Step 1.5 code from its project
• A GPU with at least 12 GB of VRAM, using CPU offload and INT8
• Python with pip

Use case | Who | What they get back
Background music | Creators | Audio generated from a text prompt
Song sketches | Musicians | A quick take on an idea
Prompt control | Developers | A guidance scale that tunes how closely it follows the text

How to set it up: Install the ACE-Step 1.5 code from the project's GitHub repository, then download this model into its checkpoints folder. The card's quick start installs the code in editable mode with pip.

• Python with pip
• The ACE-Step 1.5 project code
• About 18.8 GB of disk space for the bf16 weights

1. Install ACE-Step 1.5 as the card's quick start describes.
2. Download this model into the checkpoints folder.
3. Start the interface with the run command.

pip install -U huggingface_hub
hf download ACE-Step/acestep-v15-xl-sft

How to use it: Run the two commands below from the project folder. The first downloads the model into the checkpoints folder, and the second starts the Gradio interface with this model. Then describe the music you want and adjust the guidance scale.

huggingface-cli download ACE-Step/acestep-v15-xl-sft --local-dir ./checkpoints/acestep-v15-xl-sft
python acestep --config-path acestep-v15-xl-sft

Tool | What it does
huggingface-cli | Downloads the model into the checkpoints folder
python acestep | Starts the Gradio interface with this model
Guidance scale | Controls how closely the music follows your prompt
LM models | Optional 0.6B, 1.7B and 4B language models that all work with XL

Example: Start the interface, type a short description of the music you want, such as a slow piano piece, and generate it. Raise the guidance scale to follow your text more closely. The card says all three language model sizes work with this XL model.

For newcomers and experts: A diffusion model starts from noise and refines it over many steps until audio appears. This variant uses 50 steps for quality. A first small try is the default settings with a short prompt before you adjust anything.

For experts, the decoder has 32 layers with a hidden size of 2560 and 32 attention heads. The encoder has 8 layers at 2048. The card rates the SFT quality as higher than the base, with medium diversity against the base's high. The base variant also handles extract, lego and complete tasks.

Describe the music in text
The language model plans the composition
The diffusion decoder refines audio over 50 steps
Guidance scale keeps it close to your prompt
Listen and adjust

Good to know: The card says the model was trained on licensed music, royalty-free and public domain audio, and synthetic audio made from MIDI, and calls it ready for commercial use. Check the license terms for your own case.

• Variants: The family has a base, this SFT and a turbo version.
• Language models: The 4B one is rated stronger at audio understanding and composition.
• Memory: Full quality with the 4B language model wants 24 GB of VRAM.

Title: VRAM needed by setup in GB
CPU offload and INT8: 12 GB
CPU offload: 16 GB
No offload: 20 GB
Full quality with the 4B language model: 24 GB

Title: Inference steps by variant
Base: 50
SFT: 50
Turbo: 8

• License: MIT.
• Cost to run: none named, the source lists no API key, account or paid service, though it needs a GPU with enough memory.
• Data it reaches: only the prompts you type on your own machine, plus the files you download.
• Not for: fast drafts, since this variant takes 50 steps, or the extract, lego and complete tasks that the card lists for the base variant.

On Hugging Face, 14 November 2026

• Task: text to audio
• Library: transformers
• Parameters: 5B
• Weights: about 19.9 GB in F32; about 2.5 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 10,353
• Likes: 107
• License: mit
• Quantized versions: 1
• Fine-tunes: 7
• Tags: transformers, safetensors, acestep, feature-extraction, audio, music, text2music, text-to-audio

Links

• Model: https://huggingface.co/ACE-Step/acestep-v15-xl-sft
• Files: https://huggingface.co/ACE-Step/acestep-v15-xl-sft/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download ACE-Step/acestep-v15-xl-sft

Then follow the usage in its model card. Nothing in this kit runs by itself.