ACE-Step: Generate Music Locally in ComfyUI

ACE-Step is an open-source music generation model: give it a style description and lyrics, and it produces a full song with vocals, on your own computer. Version 1.5, released on 23 January 2026, is the one to use now, with a larger “XL” version added in April. It runs in ComfyUI or in ACE-Step’s own app, and its weights are MIT-licensed.

This guide covers which model to download for your graphics card, how to run it in ComfyUI, the settings the official workflows use, and what the licence and disclaimer mean for using the music. BFWR has not run ACE-Step; the hardware guidance below is ACE-Step’s own, and quality and speed claims are the developers’.

Quick answer

• Models: a 2B diffusion model (Turbo, SFT or Base), a 4B XL version, and a separate “planner” language model of 0.6B, 1.7B or 4B parameters.
• Memory: ACE-Step’s own table runs the 2B Turbo model on 4GB with offloading, adds the language model from 6GB, and runs XL without offloading from 20GB.
• ComfyUI: native support, with official templates. The standard one downloads about 10GB and uses 8 steps.
• Song length: 10 seconds to 10 minutes, depending on your memory.
• Licence: MIT, with a disclaimer asking you to check generated music for copyright and disclose AI use.

How ACE-Step 1.5 works

ACE-Step 1.5 has two parts. A language model acts as what the developers call “an omni-capable planner”: it turns a short request into “comprehensive song blueprints”, filling in structure, metadata and lyrics. A diffusion transformer (DiT) then generates the audio, and a VAE decodes it. The language models are fine-tuned from Qwen3 at 0.6B, 1.7B and 4B parameters; you can also skip the planner and write the tags and lyrics yourself.

The DiT comes in three versions in each size:

The XL versions use “a larger 4B-parameter DiT decoder (~9GB bf16) for higher audio quality”. Beyond text-to-music, ACE-Step lists reference-audio style guidance, cover generation, repainting parts of a song, track separation, accompaniment from a vocal track, and control of duration, BPM, key and time signature, with lyrics in “50+ languages”.

On quality, the developers say it is “between Suno v4.5 and Suno v5”, and on speed “under 10s on RTX 3090” for a full song. Both are their own claims.

Which model for your graphics card

ACE-Step’s GPU compatibility guide sets everything automatically by memory tier. A summary:

The guide notes that these tiers are “based on our testing across common hardware”, and that older NVIDIA cards below compute capability 7.0, such as the GTX 1080, fall back to a slower language-model backend automatically. ACE-Step also supports Apple Silicon, AMD ROCm and Intel GPUs, though it recommends an NVIDIA card.

Running it in ComfyUI

ComfyUI supports ACE-Step 1.5 natively. The official templates, in the Audio group, list ComfyUI 0.12.0 as the minimum:

• ACE-Step 1.5 Music Generation AIO: a single all-in-one checkpoint
• ACE-Step 1.5 Music Generation Workflow: separate files, the 2B Turbo model with the 0.6B and 1.7B language models
• ACE-Step 1.5 Music Generation (4B LLM): the same with the 4B planner
• ACE-Step 1.5XL Turbo, SFT and Base: Text to Music: the XL models
• ACE-Step 1.5 AI Lyric Song Generation: writes lyrics for you, but through a Google Gemini API node, which uses paid Comfy credits and sends your prompt off your machine. The template index marks it as not open source.

The standard workflow loads the DiT with Load Diffusion Model, both language models with a DualCLIPLoader set to type ace, and the VAE. The text node takes your style tags and lyrics separately. Its KSampler uses 8 steps, CFG 1, the euler sampler and simple scheduler, with ModelSamplingAuraFlow at shift 3, and the empty-latent node sets the length (120 seconds in the template). It saves MP3 by default.

The template examples mark lyric sections with tags such as [Intro], [Verse] and [Chorus], and write the style prompt as a list of genre, BPM, instruments and vocal style.

Files and download size

From Comfy-Org’s repackaged files, which are not gated:

The template index gives about 10.7GB for the standard workflow, 14.7GB with the 4B planner, and 19.9GB for the XL templates.

Without ComfyUI

ACE-Step’s own app is a Gradio interface that picks settings for your hardware automatically:

git clone https://github.com/ACE-Step/ACE-Step-1.5.git
cd ACE-Step-1.5
uv sync
uv run acestep

It opens at http://localhost:7860, and uv run acestep-api starts a REST API on port 8001 instead. It needs Python 3.11 or 3.12. Portable packages for Windows and macOS are linked from the README. The app also includes one-click LoRA training, which the developers say takes “8 songs, 1 hour on 3090 (12GB VRAM)” (their figure).

One caution from the project itself: its README warns about fake ACE-Step websites and says “The only official website for the ACE-Step project is our GitHub Pages site”. Download the code and models from the GitHub repository or Hugging Face.

The licence and using the music

ACE-Step 1.5’s code and model weights are MIT-licensed, which allows commercial use. Comfy-Org’s repackaged files are tagged Apache 2.0 on Hugging Face; both licences are permissive. There is no revenue limit and no content-filter clause, unlike several image models released this year.

The licence covers the software. Whether a particular song is safe to publish is a separate question, and the project’s disclaimer addresses it directly. It acknowledges “unintentional copyright infringement due to stylistic similarity” as a risk, and encourages users “to verify the originality of generated works, clearly disclose AI involvement, and obtain appropriate permissions when adapting protected styles or materials.” In practice: do not prompt for a named artist’s song, check anything you plan to release, and follow the AI-disclosure rules of the platform you publish on. This is a reading of the project’s text, not legal advice.

ACE-Step 1 or 1.5?

The original ACE-Step (v1, 3.5B parameters, April 2025, Apache 2.0) still has templates in ComfyUI, but its repository has not been updated since February 2026. Version 1.5 adds the planner language model, the Turbo model and the editing features. Use 1.5.

FAQ

Can ACE-Step run on an 8GB graphics card?

Yes. ACE-Step’s guide runs the 2B Turbo model with the 0.6B planner on 6-8GB, with offloading and INT8 quantization, for songs up to 8 minutes.

Can I use ACE-Step music commercially?

The software licence (MIT) allows it. You are still responsible for making sure a particular song does not copy existing work, and the developers ask that you disclose AI involvement.

Does ACE-Step make vocals?

Yes. Give it lyrics and it sings them, in more than 50 languages by the developers’ account. Leave the lyrics empty for instrumentals.

Does it work on a Mac?

Yes. ACE-Step supports Apple Silicon, with an MLX backend for the language model and a macOS portable package.

Related guides and tools

• Best local text-to-speech models, for speech rather than music
• ComfyUI templates: local versus API
• Local AI video generation in ComfyUI
• ComfyUI hub

Sources

• ACE-Step 1.5 on GitHub: README, model zoo, licence and GPU compatibility guide
• ACE-Step/Ace-Step1.5 on Hugging Face
• Comfy-Org ACE-Step 1.5 files
• Comfy-Org workflow templates (Audio group)

Updated 18 October 2026. Model files, templates and documentation were read on that date. Speed and quality claims are ACE-Step’s own.