ACE-Step 1.5 Review & Comparison 2026
A source-backed ACE-Step 1.5 overview: architecture deep-dive, spec-by-spec comparison with HeartMuLa and Suno, installation walkthrough, and LoRA fine-tuning notes — all drawn from official papers and repositories
ACE-Step 1.5 is the open-source model shaking up AI music generation in 2026. This guide breaks down its architecture from the official paper, compares its published specifications against HeartMuLa and Suno, walks through local installation using the official commands, and explains its LoRA fine-tuning support. Every figure below links back to a first-party source.
Understanding ACE-Step 1.5 Architecture
ACE-Step 1.5 is an open-source AI music model released under the MIT license by ACE Studio and StepFun (arXiv 2602.00744). Instead of a single end-to-end pass, it separates planning from synthesis: a language model first reasons about the song, then a diffusion transformer renders the audio. Below are the pillars of its architecture, each drawn from the official paper.
Language-Model Planner
ACE-Step 1.5 uses a language model from the Qwen3 family (0.6B, 1.7B, or 4B) as a planner. Given your prompt, it reasons step by step (chain-of-thought) to synthesize structured metadata, lyrics, and a caption that describe the target song before any audio is generated.
Diffusion-Transformer (DiT) Synthesis
The planned blueprint conditions a Diffusion Transformer (DiT) that generates the actual audio. The 2B DiT targets low-VRAM setups, while the larger XL (4B DiT) variant — released in April 2026 — raises quality at a higher memory cost.
Intrinsic Reinforcement Learning
Rather than depending on an external reward model, ACE-Step 1.5 aligns its outputs using intrinsic reinforcement learning, improving controllability and musicality directly from the model's own signals.
Open Source Stack (MIT)
The full stack — weights, code, and a Gradio UI — is MIT-licensed and hosted on GitHub and Hugging Face. ACE-Step 1.5 supports 50+ languages, LoRA fine-tuning from just a few songs, and editing workflows such as cover generation, repainting, and vocal-to-BGM.
ACE-Step 1.5 vs HeartMuLa vs Suno: Spec Comparison
An objective, source-backed comparison of published specifications — not subjective quality scores
| Property | ACE-Step 1.5 | HeartMuLa oss-3B | Suno |
|---|---|---|---|
| License | MIT | Apache 2.0 | Proprietary |
| Paper | arXiv 2602.00744 | arXiv 2601.10547 | — |
| Min VRAM | <4GB (2B) / ≥12GB (XL) | Not published | Cloud (N/A) |
| Speed | <2s (A100), <10s (RTX 3090) | RTF ≈ 1.0 | Cloud (N/A) |
| Max Duration | Up to 10 min | 240s (default) | — |
| Languages | 50+ | Multilingual | — |
| Local Deployment | Yes | Yes | No |
| LoRA Fine-tuning | Yes (few samples) | Not published | No |
| Editing | cover / repaint / vocal-to-BGM | — | — |
Sources: ACE-Step 1.5 — arXiv 2602.00744 and github.com/ACE-Step/ACE-Step-1.5; HeartMuLa — arXiv 2601.10547 and github.com/HeartMuLa/heartlib. Fields marked "Not published" are not stated in official documentation. Verified 2026-07-16.
ACE-Step 1.5 Local Installation Guide
Step-by-step guide to run ACE-Step 1.5 on your local machine, following the official repository
Check System Requirements
Pick a model to match your GPU: the 2B model runs on under 4GB of VRAM, while the XL (4B DiT) model needs at least 12GB (20GB+ recommended). The UI auto-selects the best configuration for your GPU.
Clone the Repository
git clone https://github.com/ACE-Step/ACE-Step-1.5.git && cd ACE-Step-1.5
Install Dependencies
uv sync — installs the project and its dependencies with the uv package manager.
Launch the App
uv run acestep starts the Gradio UI at http://localhost:7860. Model checkpoints download automatically from Hugging Face on first run.
Generate Music
Enter a prompt and optional lyrics in the UI and generate. To use the REST API instead, run uv run acestep-api (http://localhost:8001).
Common Issues & Fixes
CUDA Out of Memory
Switch to a smaller model (the 2B turbo variant), enable offload, or lower the quality preset. The XL model needs more VRAM than the 2B.
Gradio Port Conflict
If port 7860 is busy, set a different PORT in your .env file, or free the existing process (for example, lsof -i :7860).
Model Download Issues
Checkpoints auto-download on first run from Hugging Face or ModelScope. If a download stalls, check your network or set a custom config path via the .env file.
Windows-Specific Issues
Use WSL2 with Ubuntu for the best compatibility. A CUDA-capable NVIDIA GPU with a matching driver is required.
ACE-Step 1.5 LoRA Fine-tuning Guide
LoRA (Low-Rank Adaptation) lets you fine-tune ACE-Step 1.5 on a specific music style without retraining the full model. According to the official repository, you can train a LoRA from just a few songs by injecting small trainable matrices into the diffusion transformer — teaching the model a new genre, vocal timbre, or production aesthetic with far less compute than full fine-tuning.
Preparing Your Dataset
A well-prepared dataset matters more than a large one — a handful of clean, well-labeled tracks can be enough to shift the model's style.
- Collect a few high-quality audio samples in your target style (WAV format, 44.1kHz)
- Transcribe lyrics and tag metadata (genre, mood, tempo) for each sample
- Split into training and validation sets
Training Notes
General guidance for ACE-Step 1.5 LoRA training (tune to your data and hardware):
- LoRA rank: higher rank adds capacity but uses more VRAM
- Learning rate: start small and use a cosine scheduler
- Epochs: watch validation loss to avoid overfitting
- Batch size: scale to your VRAM (gradient accumulation helps on smaller GPUs)
HeartTranscriptor: Automated Dataset Preparation
HeartMuLa's open-source HeartTranscriptor tool helps with the most tedious part of dataset preparation. Instead of manually transcribing lyrics and tagging every audio file, it uses speech recognition and music information retrieval to generate metadata faster.
- Upload your audio files to HeartTranscriptor for automatic transcription and tagging
- Review and edit the generated metadata, lyrics, and style tags
- Export the dataset ready for LoRA training
Why Choose HeartMuLa
Production-Ready Quality
HeartMuLa delivers production-grade audio with clear vocals, consistent style, and mastered output ready for release.
Zero Setup Required
No GPU, no Python, no dependencies. HeartMuLa's cloud platform lets you generate music instantly from any browser. Sign up and create your first song in under 60 seconds.
Active Development & Support
HeartMuLa is actively developed with regular updates, a growing community, and dedicated support. Get help when you need it, not just GitHub issues.
Commercial Ready
Apache 2.0 licensed with clear commercial terms. Use generated music in any project — YouTube, podcasts, games, ads — without legal ambiguity.
ACE-Step 1.5 FAQ
What is ACE-Step 1.5?
ACE-Step 1.5 is an open-source AI music model (MIT license) that pairs a language-model planner with a diffusion transformer to create music from text prompts and lyrics. It can run locally on consumer GPUs.
Is ACE-Step 1.5 better than Suno?
They serve different needs. Suno is a polished, closed-source cloud service, while ACE-Step 1.5 is MIT-licensed and runs locally with full control. HeartMuLa is another open-source option that also offers a ready-to-use cloud platform.
How much VRAM does ACE-Step 1.5 need?
It depends on the model size: the 2B model runs on under 4GB of VRAM, while the XL (4B DiT) model needs at least 12GB (20GB+ recommended). The UI auto-selects a configuration for your GPU.
Can ACE-Step 1.5 generate vocals with lyrics?
Yes. ACE-Step 1.5 generates vocals with lyrics and supports 50+ languages, according to the official paper.
Does ACE-Step 1.5 support LoRA fine-tuning?
Yes. The official repository states you can train a LoRA from just a few songs, letting you adapt the model to specific genres or styles without full fine-tuning.
How does ACE-Step 1.5 compare to HeartMuLa?
On published specs: ACE-Step 1.5 is MIT-licensed, supports 50+ languages, and can produce loops up to ~10 minutes; HeartMuLa oss-3B is Apache 2.0. Both are open source and run locally. See the source-backed comparison table above for details.
Can I use ACE-Step 1.5 commercially?
Yes, ACE-Step 1.5 is released under the MIT license, which permits commercial use. Make sure your training data and generated content comply with copyright law in your jurisdiction.
What are the main trade-offs of ACE-Step 1.5?
The main trade-offs are the effort of local setup and the higher hardware requirements of the larger XL model. Quality and speed depend on which model size and GPU you use.
Is there a ComfyUI workflow for ACE-Step 1.5?
Community ComfyUI nodes exist for ACE-Step. For HeartMuLa, a community ComfyUI custom node (benjiyaya/HeartMuLa_ComfyUI) is available.
Should I use ACE-Step 1.5 or HeartMuLa?
Choose ACE-Step 1.5 for maximum local control, the small-VRAM 2B option, 50+ languages, and LoRA fine-tuning. Choose HeartMuLa if you prefer a hassle-free, ready-to-use cloud platform alongside an open-source model.
Experience HeartMuLa
Generate your first AI song for free — no setup, no GPU required
0/3000