ace-step/ACE-Step-1.5

ACE‑Step 1.5 – Open‑source music generation model

What it is – ACE‑Step 1.5 is a foundation model for generating full‑length music (from short loops up to 10‑minute tracks) on consumer‑grade hardware. It combines a language model that plans the song (lyrics, structure, metadata) with a Diffusion‑Transformer (DiT) that actually synthesises the audio. The repo ships the model weights, inference code, a Gradio web UI, a REST API, and tools for lightweight LoRA fine‑tuning.

Key capabilities

• Speed – <2 s per song on an A100, <10 s on an RTX 3090; can run with <4 GB VRAM for the base models.
• Quality – Benchmarked to be comparable to commercial services (between Suno v4.5 and Suno v5). XL (4 B) variants improve fidelity further.
• Control – Prompt‑driven generation with support for 50+ languages, BPM/key/scale, duration, and instrument timbres. Optional reference audio lets the model mimic a style, create covers, repaint sections, or convert vocals to accompaniment.
• Personalisation – One‑click LoRA training from as few as eight songs (≈1 h on a 3090) to capture a user’s style.
• Multi‑track & editing – Generate stems, add layers, separate tracks, and produce lyric timestamps (LRC).

How to use it

1. Install the lightweight uv package manager and run uv sync to set up the environment.
2. Launch the Gradio UI with uv run acestep (or the REST API with uv run acestep‑api).
3. Choose a DiT model (base, sft, turbo, or XL) and an optional LM (0.6 B‑4 B) – the UI auto‑selects the best configuration for your GPU.
4. Provide a textual description (or a reference audio file) and hit Generate; the system will output a wav file plus optional metadata (BPM, key, lyrics, stems, etc.).
5. For style‑specific fine‑tuning, open the LoRA Training tab in Gradio, upload a small dataset, and start training.

Model zoo – The repository hosts several DiT checkpoints:

• acestep‑v15‑base (medium quality, fast),
• acestep‑v15‑sft (higher quality, supervised‑fine‑tuned),
• acestep‑v15‑turbo (very high speed, lower CFG steps),
• XL versions (‑xl‑base, ‑xl‑sft, ‑xl‑turbo) with a 4 B DiT decoder for the best audio fidelity.

Corresponding language‑model checkpoints (acestep‑5Hz‑lm‑0.6B/1.7B/4B) provide planning, lyric generation, and audio‑understanding abilities.

Running on different hardware – The repo includes launch scripts for Windows, Linux, and macOS (Apple‑silicon via MLX). It supports CUDA, ROCm, Intel XPU, and CPU‑only modes, with automatic INT8 quantisation and off‑loading for low‑VRAM GPUs.

Community & extensions – A Discord server, an “Awesome ACE‑Step” list, and a VST3 plugin are provided for integration into DAWs. The project is MIT‑licensed, with a clear disclaimer about responsible use of generated music.

---

TL;DR – ACE‑Step 1.5 is a fast, high‑quality, open‑source music‑generation system that runs on consumer GPUs, offers extensive prompt‑based control, and includes one‑click LoRA fine‑tuning for personal style adaptation.