ACE-Step
First-generation ACE-Step model that writes full songs with vocals from tags and lyrics, fast on consumer GPUs.
About
Released in May 2025 by the ACE Studio and StepFun team, the original ACE-Step checkpoint is a 3.5B-parameter text-to-music model that writes complete songs, vocals included, from style tags plus lyrics. Rather than predicting audio tokens one by one, it runs diffusion inside a heavily compressed audio latent, which is why a minute of music renders in about two seconds on an RTX 4090 or A100. Besides plain generation it can produce variations of a take, regenerate a chosen time range, change a few lyric lines while keeping the melody, and transform an uploaded clip. The repository installs as a Python package with an acestep command that opens a local Gradio interface, and it also exposes a Python API, training scripts for full and LoRA fine-tuning, and a Chinese rap LoRA; ComfyUI nodes are available too. Plan on Python 3.10 and an NVIDIA GPU; with CPU offload enabled it fits in roughly 8 GB of VRAM, and Apple Silicon runs it at a fraction of the speed. Code and weights are Apache-2.0, so commercial use is allowed. The authors have moved on to ACE-Step 1.5, so this version mainly suits existing workflows and research comparisons.
Should you use ACE-Step?
Pick it when
Pick ACE-Step when you want full songs with vocals and lyrics on a single 8 GB GPU and turnaround matters: the authors cite up to four minutes of music in about 20 seconds on an A100, handy for drafting many song ideas in one session.
Look elsewhere when
Skip it for new projects unless you must reproduce v1 results: ACE-Step 1.5, from the same team, is the newer release with longer generation. If you want codec and lyrics transcription models alongside generation, HeartMuLa ships them.
Alternatives to ACE-Step
- ACE-Step 1.5
Newer release from the same team in its own MIT-licensed repo, with longer generation and LoRA training from a few songs, so it is the default unless you need v1.
- YuE
Multilingual 7B lyric-to-song models with in-context learning modes, but it needs 16 GB or more of VRAM and is harder to set up.
- DiffRhythm
Longer full model at about 4 minutes 45 seconds under Apache-2.0, but it fits 8 GB only with chunked decoding and is harder to set up.
- HeartMuLa
Adds a reusable 12.5 Hz codec and lyrics transcriber beside lyric-to-song generation, but its 3B model runs near real time rather than in seconds.
Reviews (0)
Leave a Review
No reviews yet. Be the first to review!
Details
- Category
- Music & Audio Generation
- Price
- Free
- Platform
- Local/Desktop
- Difficulty
- Intermediate (3/5)
- License
- Apache-2.0
- Minimum VRAM
- 8 GB
- Added
- Apr 3, 2026