ace-step

これはレビュー時に保存された抜粋です。完全で最新の内容は外部ソースを確認してください。

---
name: ace-step
description: Use ACE-Step and ACE-Step 1.5 for local or hosted AI music generation, including text-to-music, lyrics-to-song, instrumental beds, covers, repainting, stem/track extraction, track completion, LoRA personalization, REST/Python/Gradio workflows, rights review, and music integration for video, ads, games, and social content.
---

# ACE-Step music production

Use this skill when a task calls for ACE-Step or ACE-Step 1.5 as the music engine, or when choosing an open-weight/local music-generation path for songs, background music, vocal tracks, covers, remixes, stems, or video soundtrack assets.

Treat ACE-Step as a fast generative music workstation, not a deterministic MIDI sequencer. It can follow style, lyric, structure, tempo, key, and reference-audio guidance, but outputs still need listening, selection, edits, mastering, rights checks, and delivery-format prep.

## Current facts to anchor on

Documented facts below were verified on 2026-07-10 from official ACE-Step GitHub, Hugging Face model cards, official docs, and arXiv reports.

- ACE-Step 1.5 is an open-source music foundation model co-led by ACE Studio and StepFun. It combines a planning language model with a diffusion transformer acoustic renderer.
- Main generation modes include text-to-music, lyrics-to-song, instrumental generation, cover/style transfer, repainting, LEGO/layered track generation, extraction, and completion. Some editing modes are documented as base-model-only.
- The official ACE-Step 1.5 Hugging Face card lists license `MIT`, model type `Text2Music`, 50+ languages, and consumer-hardware/local operation claims. The GitHub repository is MIT-licensed.
- The official API is asynchronous: submit with `POST /release_task`, poll `POST /query_result`, then download files via `/v1/audio?path=...`.
- Duration control exists through `audio_duration` / `duration` / `target_duration`; the official API docs list range 10-600 seconds. The technical report describes v1.5 as scaling from short loops to 10-minute compositions.
- Output format options documented in the API include `flac`, `mp3`, `opus`, `aac`, `wav`, and `wav32`.
- Official model-zoo guidance distinguishes DiT model families:
  - `acestep-v15-turbo`: fast, SFT, 8-step, high quality, medium diversity, medium fine-tunability; good first default for production iteration.
  - `acestep-v15-base`: 50-step, CFG-enabled, medium quality, high diversity, easy fine-tuning; use when you need base-only tasks, CFG, higher exploration, extraction/LEGO/complete support, or training/fine-tuning workflows.
  - `acestep-v15-sft`: 50-step, high quality, medium diversity, easy fine-tuning; use when you prefer SFT quality over turbo speed and do not need base-only capabilities.
  - XL models use a larger 4B DiT decoder; official docs say they target higher quality and require more VRAM, with >=12GB VRAM using offload/quantization or >=20GB without offload.
  - LM options include `acestep-5Hz-lm-0.6B`, `1.7B`, and `4B`; larger LMs are documented as stronger at composition and melody copying.
- Official docs state launch scripts exist for Windows CUDA, Windows ROCm, Linux CUDA, and macOS Apple Silicon MLX. Custom launch settings can be kept in `.env` using variables such as `ACESTEP_CONFIG_PATH`, `ACESTEP_LM_MODEL_PATH`, `PORT`, and `LANGUAGE`.
- ACE-Step DAW is a separate AGPL-3.0-or-later project and has different distribution obligations than the MIT model repository; do not assume MIT terms apply to a bundled DAW deployment.

Sources:

- Official repository: https://github.com/ace-step/ACE-Step-1.5
- Official model card: https://huggingface.co/ACE-Step/Ace-Step1.5
- Official API docs: https://github.com/ace-step/ACE-Step-1.5/blob/main/docs/en/API.md
- Official inference docs: https://github.com/ace-step/ACE-Step-1.5/blob/main/docs/en/INFERENCE.md
- Official tutorial: https://github.com/ace-step/ACE-Step-1.5/blob/main/docs/en/Tutorial.md
- Official LoRA training tutorial: https://github.com/ac