GitHub - ace-step/ACE-Step-1.5: The most powerful local music generation model that outperforms almost all commercial alternatives, supporting Mac, AMD, Intel, and CUDA devices.

The most powerful local music generation model that outperforms almost all commercial alternatives, supporting Mac, AMD, Intel, and CUDA devices. - ace-step/ACE-Step-1.5

Summary

ACE-Step 1.5 is an open-source text-to-music foundation model that pairs a Language Model (acting as a planner) with a Diffusion Transformer (DiT) decoder to generate songs locally on consumer hardware. It ships with multiple interfaces—a Gradio UI, REST API, Python API, CLI, and a standalone VST3 plugin—and supports CUDA, Apple MPS, AMD ROCm, Intel XPU, and CPU backends. The project is MIT-licensed and written primarily in Python.

What it is useful for

ACE-Step 1.5 covers both generation and editing workflows. From a text prompt, lyrics, or simple description, the model produces audio ranging from 10 seconds up to 10 minutes, and can run up to 8 songs in parallel as a batch. Beyond text-to-music, the project documents cover generation, repaint (local audio editing), vocal-to-BGM conversion, track separation, and multi-track layering.

Engineers can also use it for lightweight personalization: a LoRA can be trained from as few as 8 songs in roughly an hour on an RTX 3090 with 12GB VRAM, capturing a specific style. Metadata control is exposed for duration, BPM, key/scale, and time signature, and the model supports lyrics in 50+ languages.

How engineers can use it

The documented installation path uses the uv toolchain and Python 3.11–3.12:

curl -LsSf https://astral.sh/uv/install.sh | sh
git clone https://github.com/ACE-Step/ACE-Step-1.5.git
cd ACE-Step-1.5
uv sync
uv run acestep        # Gradio UI on http://localhost:7860
uv run acestep-api    # REST API on http://localhost:8001

Platform-specific launch scripts (start_gradio_ui.sh, start_gradio_ui.bat, start_gradio_ui_macos.sh, and ROCm variants) are included, and a .env.example file lets you override the model, port, and language without editing the repo.

Model choice depends on available VRAM. The README's tier table recommends the 2B turbo DiT plus a 0.6B LM for 6–8GB GPUs, a 2B DiT with a 1.7B LM on vLLM for 8–16GB, and the 4B XL DiT (with the 4B LM) for ≥24GB cards. XL models require ≥12GB VRAM with offload and quantization, or ≥20GB without.

Documented limitations to plan around: ROCm on Windows is supported only with Python 3.12 (the wheel AMD provides officially); XL models cannot fit on sub-12GB cards even with offload in some configurations; and the project's own disclaimer flags risks around stylistic similarity to copyrighted material, urging verification of originality. Full per-platform instructions live in docs/en/INSTALL.md.