Local music generation
Run ACE-Step 1.5 locally: setup through your first song
Write lyrics, describe the genre and instruments you want, then generate a first song on your own computer. This guide follows the official ACE-Step 1.5 repository setup, opens the local interface, checks pronunciation and chorus structure with short lyrics, and shows how to refine a take by recording duration, seed and settings. The install baseline is Python 3.11 or 3.12 and about 10 GB of storage; standard 2B and XL 4B models have different NVIDIA memory needs. Apple-silicon Macs have a dedicated MLX1 launcher. Published speed figures omit song length and detailed conditions, so they cannot stand in for a result on your device.
ACE-Step and ACE Studio are different products
ACE-Step 1.5 is a model project whose public code and weights you run yourself. ACE Studio is commercial production software with AI singers, instruments, vocal editing and generation. Its docs describe product models such as Verse25 and features using AI Credits. Choose ACE-Step to handle the model directly, or ACE Studio for an integrated workstation.
The documented UI languages include English, Chinese, Hebrew and Japanese. The model card advertises more than 50 languages, not equal quality in each. Test Korean pronunciation and syllables with a short passage.
Choose model size for your hardware
The official install guide lists Python 3.11–3.12 and about 10 GB disk. For the standard 2B DiT2, starting VRAM3 guidance is about 4 GB for DiT-only or 6 GB with LM+DiT. XL is a separate 4B DiT: the README distinguishes 12 GB or more with offload/quantization from 20 GB or more without offload4. Reserve additional memory for other apps.
The current NVIDIA table recommends 2B turbo DiT-only at 6 GB or less; 0.6B LM at 6–8 GB; 0.6B/1.7B LM at 8–16 GB; 2B SFT or XL turbo at 16–20 GB (XL needs CPU5 offload); XL turbo/SFT at 20–24 GB (no offload, 1.7B LM, 4B LM available); and XL SFT plus 4B LM at 24 GB or more. Do not conflate XL with standard 2B.
These are UI recommendations; consult the GPU6 Compatibility Guide for duration and batch limits.
On Apple silicon, the dedicated start_gradio_ui_macos.sh configures the MLX backend. The install guide lists M1/M2/M3/M4; this is a different runtime7 path from the NVIDIA VRAM table. CPU mode is possible but documented as much slower.
Start the local UI from the official repository
The official Quick Start clones the repository, syncs dependencies with uv, then launches the UI. First run may download model weights and need extra time and disk space. The documented UI binds to 127.0.0.1:7860 by default. Keep this for personal use and leave public sharing off.
git clone https://github.com/ACE-Step/ACE-Step-1.5.git
cd ACE-Step-1.5
uv sync
uv run acestepSeparate lyrics and tags for your first song
Put genre, mood, instruments and vocal character in the prompt, and sung words in the Lyrics field. Fields and options can change by version, so follow the UI you have. Describe audible traits, such as bright indie pop, warm electric piano, soft vocals and drums that lift in the chorus.
Section labels such as [Verse] and [Chorus] may not be followed exactly. Start with a short verse and chorus to check pronunciation, repetition and form; extend duration gradually. The model card's up-to-ten-minute description does not guarantee practical ten-minute output on every device.
Record seeds and settings when revising
If the UI exposes a seed, save it for comparisons under similar conditions. A different model or code version can still change the output. If no seed option is visible, do not invent CLI8 flags; save the prompt and settings.
Change one thing at a time: shorten lyrics when diction is unclear, or reduce instrument instructions when the arrangement is crowded. Save the model, duration, seed and prompt with a useful take for easier comparison.
Check published speed claims on your device
The model card claims under two seconds on an A100, under ten seconds on an RTX 3090 and operation below 4 GB VRAM. It omits song duration, settings, precision, warm-up, measurement code and an actual VRAM trace. These are not controlled cross-model benchmarks or user guarantees.
On your device, generate the same short lyrics, tags and duration several times; record model loading separately from generation. Log memory, output length and whether you used MPS or CUDA9. No official Apple-silicon generation time was found.
Check licenses and source material
The ACE-Step 1.5 model card lists MIT. Code and weights can be distributed separately, so check the LICENSE and model card for your downloaded version. This does not replace ACE Studio terms or rights in supplied lyrics and reference audio.
Before publishing or commercial use, check rights in supplied lyrics, vocal references and prompts naming specific artists. Local inference10 itself does not grant unrestricted output rights.
Terminology notes
MLX — A machine-learning framework for Apple silicon. It uses Apple silicon’s unified-memory architecture; supported models and features vary by MLX tool.
Back to the textDiT — Short for Diffusion Transformer: a model architecture that uses a transformer to refine noisy representations across diffusion steps.
Back to the textVRAM — Memory used by a graphics card’s GPU for model weights and intermediate values. It is distinct from system RAM.
Back to the textOffloading — Moving some model data from GPU memory to system RAM or storage when capacity is limited. This adds data transfer.
Back to the textCPU — The central processor that executes general-purpose program instructions. AI workloads may divide work between it and other processors such as GPUs.
Back to the textGPU — A processor designed to handle many calculations in parallel. It performs model computations during AI inference.
Back to the textRuntime — Software that loads model files and runs their computations. Supported formats, hardware, and optimizations vary by runtime.
Back to the textCLI — Short for Command-Line Interface: operating a program by entering commands in a terminal.
Back to the textCUDA — A software platform for general-purpose computing on NVIDIA GPUs. Programs built for CUDA are not guaranteed to run unchanged on other GPUs.
Back to the textInference — The process of using a trained model to compute an output for an input. Here, local inference means running the model on the user’s device.
Back to the text