Meet ACE-Step 1.5, an on-device music model that beats Suno on common eval metrics

ACE-Step 1.5 is an open-source music model that can generate a full song in about 2 seconds on an A100, runs locally on a typical PC (around 4GB VRAM), and beats Suno on common evaluation scores.

GitHub: https://github.com/ace-step/ACE-Step-1.5

Since SUNO arrived, AI music has swept across the world. But in the music model landscape, there's still a big gap between open-source and closed-source models, whether in generation quality, capability diversity, or raw performance.

Right now, this space is dominated by commercial closed-source models. That means our "creative power" is often tied to a specific app and a specific model. If you lose access to that model, or the model changes in ways you didn't choose, your ability to create can be impacted overnight (you can see this kind of concern in community discussions, for example:https://www.reddit.com/r/SunoAI/comments/1p76vo0/rip_suno_youre_going_to_to_be_missed/.

With ACE-Step 1.5, we want to break that dependency by giving the world a truly competitive open-source option. Creators should be able to run the model on their own computer, own it, fine-tune it with their own songs, and worry less about privacy and data leakage.

What is ACE-Step 1.5

Key Traits:
It's truly fast. On a single A100, generating a complete 3-minute song takes about 2 seconds. On consumer hardware, it still runs at an acceptable speed (around 10 seconds on an RTX 3090, and it takes about 4GB VRAM).

It delivers high generation quality. On common evaluation metrics, it beats SUNO and other commercial music models, and it creates a clear gap compared to existing open-source music models.

It supports user LoRA fine-tuning. With just a few songs, you can train your own LoRA and capture your style.

It comes with broad, practical capabilities, including:

Full song generation

Song cover

Repainting

Add a stem based on the existing track

Generative stem splitting

Song completion

What you can do with ACE-Step 1.5:

Because it's fast and runs locally, ACE-Step 1.5 unlocks a lot of new workflows.
Batch generation plus automatic scoring:

Since it runs locally and is fast enough, you can generate 100 variations from the same prompt and then pick the best one. To support this, we're also open-sourcing a reward model that can score the generations. You can let the model generate 100 songs, then simply choose the one with the highest score to get a strong result quickly.

LoRA fine-tuning:

With a few songs, or a few dozen songs, you can train your own LoRA. In our experiments, we trained a LoRA using 8 Chinese New Year (Spring Festival) style songs. After that, when we generated totally different genres like love songs, jazz, and rock, they all carried a strong Chinese New Year vibe. LoRA lets creators fine-tune toward a specific style using their own music. It learns from your songs and captures your style. And because you run it locally, you own the LoRA and don't have to worry as much about data leakage.

Cover:

Feed in any song, then provide a prompt plus lyrics, and you can transform the track into a different style. We've tried turning a vocal-forward pop song into a synth-driven instrumental dance track, and turning a soft female vocal song into a male-voice rock version. This differs slightly from SUNO's cover feature: it keeps the input song's structure and duration, but reimagines the melody and instrumentation. What you build with it is up to your creativity, and we can't wait to see what you do.

Repaint:

This is a super handy feature. You load a song, select a segment, click "Repaint," and regenerate that segment while keeping the audio before and after it intact. If you don't like a small part of an otherwise great track, you can keep repainting that section. And yes, ACE-Step is fast, so feel free to repaint 100 times to land on the one you love.

A fun community app: Vibe DJ:

Leveraging ACE-Step's local runtime, high speed, and Repaint capability, a community developer built a very cool app called VibeDJ. Each time you type a prompt, the model generates a fixed-length segment and plays it. While it's playing, you type the next prompt, then use Repaint to generate the next segment so it connects seamlessly with the previous one.
This turns anyone into a DJ using natural language prompts. You "DJ" by typing instead of mixing. Imagine pulling out your laptop at a house party and running a live VibeDJ set for everyone. That's the kind of thing local, real-time generation makes possible.

ComfyUI Integration

We're excited to share that ACE-Step 1.5 is already available in ComfyUI, thanks to their day-0 support 🙌

Tutorial (how to use ACE-Step 1.5 in ComfyUI):https://blog.comfy.org/p/ace-step-15-is-now-available-in-comfyui

With ACE-Step 1.5, the only limit is your imagination

Explore the ACE-Step repository here: https://github.com/ace-step/ACE-Step-1.5
If you enjoy ACE-Step 1.5, we'd appreciate it if you could star the repo:

acemusic.ai

Based on ACE-Step 1.5, we're building an open ecosystem AI music platform: acemusic.ai.

On acemusic.ai, we provide a permanently free cloud version (truly free, with no daily limits), along with local installation and deployment packages for running the model yourself. We will never restrict downloads or uploads, we support LoRA sharing, and we plan to integrate more great open-source models as well as third-party music model APIs. Our goal is to offer an open, full-featured AI music platform, and return creative freedom, asset ownership, and model ownership back to creators.
Feel free to give it a try: acemusic.ai

- - -
You can find more about ACE-Step 1.5 on:
Project page: https://ace-step.github.io/ace-step-v1.5.github.io/
Discord:https://discord.gg/aMjaXPUg7U
X:https://x.com/acemusicAI
YouTube: https://www.youtube.com/@acemusicAI
ACE Music: https://acemusic.ai/

5655 Radford Avenue
Valley Village, CA 91607
United States

ACE-Step 1.5 is an open-source music model that can generate a full song in about 2 seconds on an A100, runs locally on a typical PC (around 4GB VRAM), and beats Suno on common evaluation scores.

This release was published on openPR.