ACE-Step Guide: Free, Open AI Music You Actually Own

Most AI music tools rent you a sound. Suno’s free plan keeps the rights to everything you generate, Udio switched off downloads in October 2025 and hasn’t turned them back on, and both companies are still in court with major labels. ACE-Step changes the deal: you download it, run it on your own machine, and nobody upstream gets a say in what you do with the output.

Here’s the short version of this ACE-Step guide. ACE-Step 1.5 is a free, open-source music generation model from ACE Studio and StepFun, released under the MIT license, that runs on a consumer graphics card and renders full songs with vocals in seconds. The license lets you use the output commercially with no subscription and no monthly download cap. What it can’t do is hand you a copyright. In the United States, a track the model made on its own isn’t copyrightable, so “you actually own it” really means nobody else has a claim on it either, and the parts you add yourself are the parts you can register and defend.

What ACE-Step Is and Where It Came From

ACE-Step is a joint project between ACE Studio and StepFun. The first version landed in May 2025 as a 3.5 billion parameter model under the Apache 2.0 license, and its paper set out a goal that stuck with me: a stable diffusion moment for music, meaning a fast, open base model that other people could build on rather than another closed app.

Version 1.5 arrived on January 28, 2026, and it’s the one worth your time. A small language model now acts as a planner: it takes your rough description, works out structure, tempo, key, and lyrics if you didn’t supply any, and hands that blueprint to a diffusion transformer, the part that actually renders audio. The ACE-Step 1.5 repository on GitHub lists generation times of under two seconds for a full song on an A100 and under ten seconds on an RTX 3090, with a minimum footprint of less than 4GB of video memory. You can generate anything from a ten second loop to a ten minute piece, and lyrics work across more than 50 languages.

On April 2, 2026 the team added an XL line with a 4 billion parameter decoder for better audio quality, in three flavors: base, sft, and turbo. The code and weights for 1.5 are under the MIT license, which is about as permissive as software licenses get. The repository showed 12,800 stars when I checked, and the Hugging Face model card reported just under 70,000 downloads in the previous month.

The feature list is longer than most paid tools can offer: reference audio to guide style, cover generation, repainting a chosen section without touching the rest, stem separation, turning a vocal into a full backing track, and explicit control of BPM, key, and time signature. There’s also one-click LoRA training inside the web interface. LoRA, short for low rank adaptation, is a small add-on file that teaches the model a style without retraining the whole thing, and the README says eight songs and about an hour on a 3090 is enough to make one.

The Ownership Question, Answered Plainly

This is the part people get wrong, so I’ll separate two things that usually get mashed together: the license and the copyright.

The license is the easy half. MIT means you can run the model, modify it, build a product on it, and sell what it makes. There’s no free tier that quietly keeps your songs, no commercial rights toggle that costs $10 a month, and no clause that lets a company revoke your use of a track you made last year. According to the ACE-Step 1.5 model card on Hugging Face, the training set was built from licensed tracks, royalty-free and public domain recordings, and synthetic audio rendered from MIDI, and the card states outright that generated music can be used for commercial purposes. I can’t audit that dataset and neither can you, but it’s a very different posture from the companies being sued for training on major label catalogs.

The copyright is the hard half, and the license has nothing to do with it. The US Copyright Office’s Part 2 report on copyrightability, published January 29, 2025, concluded that material generated entirely by AI isn’t protected, that prompting alone doesn’t make you the author of the output, and that the human-authored parts of a mixed work can be registered as long as the AI parts are disclosed and excluded from the claim. As Celeste Moy of Rimon Law explains in her analysis of how the Office’s guidance applies to music, the Office had already registered more than a thousand works where applicants followed that disclosure route. The Supreme Court declined to revisit the human authorship requirement when it denied certiorari in Thaler v. Perlmutter on March 2, 2026, so the rule is settled for now.

So what does “you actually own” mean in practice? The output is clean of anyone else’s claim. No platform owns it, no label has licensed it, and the model’s training story is defensible. What you don’t get is exclusivity over the raw generation. If someone copies a purely AI-generated ACE-Step track, you have no copyright to enforce. Real ownership comes from putting real authorship into it, which I’ll come back to below.

Two regional notes. In the UK, the government’s report on copyright and AI, published March 18, 2026, proposed repealing the old provision that gave computer-generated works with no human author a 50 year term, while keeping protection for AI-assisted works, and Mishcon de Reya points out that a sound recording would still be protected there even if the composition isn’t. In the EU, Article 50 of the AI Act has applied since August 2, 2026, and I’ll cover its labeling duties in the distribution section.

Where It Beats Suno and Udio, and Where It Doesn’t

I’ve paid for both Suno and Udio, and I still keep a Suno subscription for comparison. Here’s how the three sit in October 2026.

The Suno numbers come from Suno’s own pricing page: Pro drops to $8 a month and Premier to $24 a month on annual billing, the free plan gets 50 credits a day with no downloads and no commercial rights, and the paid tiers now cap song downloads at 20 and 60 per month. That cap matters if, like me, you render 30 variations of a chorus before liking one.

Udio is the cautionary tale. It settled with Universal on October 29, 2025, switched off downloads the next day, opened a 48 hour export window in early November, and then closed the gate. According to the review published by AI Tools Directory on October 4, 2026, downloads still haven’t returned and the licensed platform Udio promised hasn’t launched. You can pay $30 a month and never hold a file.

On raw sound, I want to be fair to Suno. The ACE-Step team’s own README places 1.5 between Suno v4.5 and v5, and the independent review at Z.Tools lines up with that: ACE-Step 1.5 scored 8.09 on the SongEval benchmark against 7.87 for Suno v5, but Suno won on the AudioBox audio quality measure, 7.87 to 7.42, and human listeners in the paper placed ACE-Step between those two Suno generations. In my sessions that means convincing structure, accurate tempo and key when you specify them, and the occasional vocal with a metallic edge or a smeared consonant. Suno’s mixes are more polished out of the box. ACE-Step’s are more useful, because I can open the stems and fix things.

Where ACE-Step wins outright is control. Seeds are deterministic, so I can regenerate the same song with one change, or repaint bars 33 through 48 and leave the rest alone. I can feed it a demo from my phone as reference audio and have it build around my own melody. And I can train it on my own back catalog, which no paid tool allows at this level of access. For local AI music generation you control end to end, nothing else comes close right now.

What You Need to Run It, and What It Costs

Almost anything with a dedicated GPU will run the small models, and the good stuff starts at 12GB of VRAM. VRAM is the memory on your graphics card, and it’s the number that decides which models you can load.

The project’s own tiering is worth following. With 6GB or less you run the 2B turbo decoder with no planner, INT8 quantization, and full CPU offload, which works but means you write the full prompt and lyrics yourself. From 6 to 8GB you add the 0.6B planner. From 8 to 16GB you get the default setup of turbo plus the 1.7B planner, and that’s where I’d point most people. Between 16 and 20GB you can run the higher quality sft decoder or XL turbo with offload. At 20GB and up the XL models fit without offload, and at 24GB you can pair XL sft with the 4B planner for the best results the project offers.

Turbo versus sft is the other decision. Turbo renders in 8 diffusion steps and is what you’ll use for 90 percent of iteration. Sft uses 50 steps, takes several times longer, and produces a cleaner result; I switch to it for the final render once I’ve locked a seed I like. The base model exists for the extract, lego, and complete editing tasks and isn’t the pick for ordinary text-to-music work.

Beyond the GPU you need Python 3.11 or 3.12, roughly 10GB of disk for the core models, and patience for the first download. The installation guide covers Windows, Linux, and macOS, and Apple Silicon is a first-class target: the macOS launch scripts run the planner through Apple’s MLX framework, with unified memory standing in for VRAM. AMD cards work through ROCm, Intel GPUs through XPU, and a CPU mode exists for inference only. Windows and Mac users who don’t want to touch a terminal can grab the portable packages linked from the README.

If you have no GPU at all, three routes. The README points to acemusic.ai as a free hosted version. The Hugging Face Space demo works but you’ll share a queue. Or rent a card by the hour: RunPod’s pricing page listed an RTX 4090 from $0.34 an hour as of its July 2026 update, though its rates have moved up several times this year, so check before you commit. A 4090 renders a full ACE-Step song in seconds, so an hour of rental buys a long session.

One warning from the maintainers’ own README: the project says it operates no websites beyond its GitHub pages, names several lookalike domains built around the model’s name, and asks people not to pay them. Credits for a model that costs nothing are a scam.

For DAW people, two integrations matter. ComfyUI has supported ACE-Step since version 1, with a documented workflow on the ComfyUI docs site. And the team now ships acestep.vst3, an official VST3 instrument built on a C++ port of the model that runs on CPU, CUDA, Metal, or Vulkan and loads in Reaper or Ableton, so you can generate directly onto a track.

How I Actually Work With It

My setup is a 16GB card, turbo plus the 1.7B planner for drafts, and the following loop.

I start in custom mode, not simple mode. Simple mode is fine for a demo, but it lets the planner write lyrics and structure, and anything the planner writes is AI-authored material you can’t claim. I paste my own lyrics with section tags, set BPM, key, and time signature by hand, and give a style description in plain language rather than a comma soup of tags. The model accepts tags, descriptive text, or use-case scenarios, and in my experience two or three descriptive sentences about instrumentation, era, and mood beat a list of 15 genre words.

I render in batches of four to eight and audition at low volume on headphones, listening for structure and feel rather than polish. When one lands, I note the seed, then switch to repaint for the sections that don’t work. Repaint is the single most useful thing in the tool: it regenerates a chosen time range while keeping the rest, so a flat second verse doesn’t cost you a great chorus. The one place I’ve been burned is transitions at repaint boundaries, which can sound stitched; the Z.Tools review flags the same thing, and my fix is to overlap the repaint range by a bar on each side and crossfade in the DAW.

For a final pass I re-render the locked seed with the sft decoder, run stem separation, and bring the stems into Logic. From there it’s ordinary production: replace the AI vocal with my own take or a session singer, re-amp guitars, swap the drums, and mix. Reference audio I only ever feed from my own recordings. Pointing the cover or reference modes at someone else’s released track makes a derivative work of something you don’t own, and it defeats the point of using a clean model.

Turning a Generation Into a Record You Can Defend

Here’s where ownership gets built, and the model can’t do it for you. The Copyright Office protects human expression, so the question for every release is: which parts did a person make? Lyrics you wrote are yours. A melody you sang or played and then had the model arrange around is yours. Vocals you tracked, parts you re-recorded, and the arrangement decisions you made in the DAW all count. The AI-rendered instrumental, left as generated, does not.

When you register, Reed Smith’s guide to music and AI puts it bluntly: applications should accurately identify the human authorship and exclude more than trivial AI material from the claim. In practice that means using the Standard Application, describing your contribution in the New Material Added field, and listing the AI-generated portions in Material Excluded. The musicologize.com guide to registering AI-assisted songs adds a point I’d underline: if you previously registered a track without disclosing AI content, the Office expects a supplementary registration to fix the record.

Timing matters this fall. As Hypebot reported in September 2026, the Office’s new fee schedule takes effect in mid-November 2026. The Standard Application goes from $65 to $85, the Single Application from $45 to $55, and the Group Registration for Works on an Album of Music, the option that files a whole album as one claim, doubles from $65 to $130. If you have a batch of finished songs with real human authorship in them, file before the change.

Performing rights organizations have caught up too. Since October 28, 2025, ASCAP, BMI, and SOCAN have all accepted registrations for partially AI-generated works, meaning compositions that combine AI-generated content with human authorship, while works that are entirely AI remain ineligible. Write the lyrics and the hook yourself and you have a registrable composition. Prompt a whole song and you don’t.

If a track matters to you and the AI parts are load bearing, the cleanest route I know is to have real musicians re-record it using the generation as a guide, so the arrangement, parts, and performances are made by people. That gives the recording a clear chain of human authorship, and it’s what I do for anything I’d put a marketing budget behind. It costs more than a click. It’s also the only version of “I own this” that survives contact with a lawyer.

What Distributors and Streaming Services Will Accept in 2026

Policy moved fast this year, and the distributor choice is now the first decision, not the last.

DistroKid is the clearest yes among the big three self-serve services. Its support article on music made with AI tools says AI-made music is accepted provided you own 100 percent of the rights, don’t imitate anyone’s voice or identity, don’t mass-upload generic content to game the algorithms, and don’t infringe anything, and it warns that stores can still reject or pull a release. TuneCore’s public position is that it won’t distribute works that are 100 percent AI-generated but supports AI that enhances human creation. CD Baby is the hard no: its support language, as reviewed by Velveteen in July 2026, says it cannot accept AI-generated content even if commercially licensed.

Spotify doesn’t ban AI music and said so plainly in its September 25, 2025 policy announcement, which laid out three measures: a firmer impersonation policy covering unauthorized voice clones, a spam filter that tags and stops recommending mass uploads, duplicates, and very short tracks, and support for a DDEX credits standard so artists can state where AI was used. DDEX is the metadata standard distributors use to send track information to stores. The same post reported over 75 million spam tracks removed in the preceding year, and a later update added that AI credits went into beta on April 16, 2026, with tens of thousands submitted daily. Spotify has also said profiles that mainly represent AI persona artists aren’t eligible for its verification badge.

The scale problem explains the tone. According to Music Business Worldwide’s explainer on the EU AI Act, Deezer said fully AI-generated music made up more than half of all new uploads to its platform in June 2026, nearly 90,000 such tracks a day, and Deezer tags them at the platform level. TIDAL set a policy in June 2026 to tag tracks it identifies as fully AI-generated and stop them earning royalties. Apple Music launched tags in March 2026 that depend on labels and distributors declaring AI content, and Amazon Music doesn’t flag it at all.

For readers in the EU, the same piece explains Article 50: since August 2, 2026, providers of generative systems must mark output so it’s machine-detectable, with a grace period to December 2 for systems already on the market, and anyone publishing deepfakes, which includes audio cloning a real artist’s voice, must label them clearly, with fines up to 15 million euros or 3 percent of turnover. An open model you run yourself puts that burden on you.

My rule across all of this: disclose. Tick the AI box at your distributor, fill in the DDEX fields, and build human parts into the track so the disclosure reads as AI-assisted rather than fully AI-generated. The platforms are separating those two categories, and the fully generated one is being tagged, demonetized, and filtered. That’s the AI music copyright and distribution reality in 2026, and no license changes it.

An Honest Reality Check

ACE-Step is the best open-source AI music generator I’ve used and it’s still not a finished product. The vocals are the weak link; I get a usable lead maybe one take in four on English pop, worse on rap, and the team’s own notes admit rough vocals and weak results in some genres. The default mixes sit a little dull and need a mastering pass. Output varies run to run in ways Suno has mostly ironed out, which is why the batch and seed workflow matters.

It’s also software you maintain. There’s no support desk, the repository shows more than 1,400 commits, and a driver change can break your setup on a Tuesday. LoRA training is advertised at 12GB but the training tutorial recommends 20GB or more for full length songs, and training on anyone’s catalog but your own is a legal mess. Finally, remember what the license can and can’t do. MIT gives you freedom to use the output. It does not give you a copyright, and no open license ever will.

What to Do This Week

Pick the path that matches your hardware. With a 12GB or larger NVIDIA card, clone the repo, run the installer, and let it auto-select turbo plus the 1.7B planner. On an Apple Silicon Mac with 16GB or more of unified memory, grab the macOS portable package. With neither, open acemusic.ai or rent a 4090 for an hour.

Then write a lyric, set a tempo and key, and generate eight takes of a single chorus. Pick one, note the seed, repaint whatever bothers you, and render the winner with the sft decoder. Pull the stems into your DAW and replace at least the lead vocal with a human one. Tick the AI disclosure at your distributor, and if the song has real authorship in it, file a Standard Application with the Copyright Office before the fees rise in November. That’s an ACE-Step song you can release, register, and defend, and it cost you nothing but the work.

Sources

• GitHub, ACE-Step 1.5: Pushing the Boundaries of Open-Source Music Generation: https://github.com/ace-step/ACE-Step-1.5
• Hugging Face, ACE-Step/Ace-Step1.5 Model Card: https://huggingface.co/ACE-Step/Ace-Step1.5
• arXiv, ACE-Step 1.5: Pushing the Boundaries of Open-Source Music Generation: https://arxiv.org/abs/2602.00744
• arXiv, ACE-Step: A Step Towards Music Generation Foundation Model: https://arxiv.org/abs/2506.00045
• Hugging Face, ACE-Step/ACE-Step-v1-3.5B Model Card: https://huggingface.co/ACE-Step/ACE-Step-v1-3.5B
• GitHub, ACE-Step 1.5 Installation Guide: https://github.com/ace-step/ACE-Step-1.5/blob/main/docs/en/INSTALL.md
• GitHub, ACE-Step 1.5 Ultimate Guide: https://github.com/ace-step/ACE-Step-1.5/blob/main/docs/en/Tutorial.md
• GitHub, ACE-Step 1.5 LoRA Training Tutorial: https://github.com/ace-step/ACE-Step-1.5/blob/main/docs/en/LoRA_Training_Tutorial.md
• GitHub, acestep.vst3: https://github.com/ace-step/acestep.vst3
• ComfyUI Docs, ACE-Step ComfyUI Text-to-Audio Generation Workflow Example: https://docs.comfy.org/tutorials/audio/ace-step
• Z.Tools, ACE-Step v1.5 review: open-source AI music after Suno v5: https://z.tools/blog/ace-step-v1-5-review
• Suno, Pricing: https://suno.com/pricing
• AI Tools Directory, Udio Review 2026: Licensed and Capable, But You Can’t Download Your Songs: https://www.aitools-directory.com/blog/udio-ai-review-2026/
• Rimon Law, How Copyright Office Guidance Applies to Music That Includes AI-generated Material: https://www.rimonlaw.com/how-copyright-office-guidance-applies-to-music-that-includes-ai-generated-material/
• Reed Smith, Music Part 2: Copyright and NIL: https://www.reedsmith.com/articles/entertainment-media-guide-to-ai-three-years-on/music-part-2-copyright-nil/
• Musicologize, Can I Copyright a Song I Made with AI?: https://www.musicologize.com/can-i-copyright-a-song-i-made-with-ai/
• Hypebot, 2 Months Left Before U.S. Copyright Fees Jump Dramatically: https://www.hypebot.com/2-months-left-before-u-s-copyright-fees-jump-dramatically/
• BMI, ASCAP, BMI and SOCAN Announce Alignment on AI Registration Policies: https://www.bmi.com/news/entry/ascap-bmi-and-socan-announce-alignment-on-ai-registration-policies
• DistroKid Help Center, Can I Upload Music Made With AI Tools to DistroKid?: https://support.distrokid.com/hc/en-us/articles/41182362733715-Can-I-Upload-Music-Made-With-AI-Tools-to-DistroKid
• Velveteen, Which Music Distributors Accept AI Music in 2026: https://www.velveteen.fm/guides/ai-tools-for-independent-artists/which-music-distributors-accept-ai-music
• Spotify Newsroom, Spotify Strengthens AI Protections for Artists, Songwriters, and Producers: https://newsroom.spotify.com/2025-09-25/spotify-strengthens-ai-protections/
• Music Business Worldwide, The EU AI Act’s rules on AI content labeling and training are now law and enforceable. Here’s what they mean for the music business: https://www.musicbusinessworldwide.com/the-eu-ai-acts-rules-on-ai-content-labeling-and-training-are-now-law-heres-what-they-mean-for-the-music-business/
• Stephenson Harwood, UK government publishes report and impact assessment on Copyright and Artificial Intelligence: https://www.stephensonharwood.com/post/102mpke/uk-government-publishes-report-and-impact-assessment-on-copyright-and-artificial-intelligence-the-debate-continues/
• Mishcon de Reya, Retail Horizon Scan July 2026: https://www.mishcon.com/reports/retail-horizon-scan-july-2026/page-11
• RunPod, GPU Cloud Pricing: https://www.runpod.io/gpu-cloud/pricing