README.md

March 8, 2026 · View on GitHub

🎛️ ProdIA-MAX

Enhanced fork of ACE-Step UI — AI Music Production Suite for Windows
Fork mejorado de ACE-Step UI — Suite de producción musical con IA para Windows

Version MIT License Windows ACE-Step 1.5

ProdIA-MAX Screenshot


🙏 Credits / Créditos

ProdIA-MAX is a fork and extension of ACE-Step UI by fspecii.
All original UI code, architecture, and design belong to their respective authors.
This project adds Windows-specific tooling and production-focused enhancements on top of their work.

ProdIA-MAX es un fork y extensión de ACE-Step UI creado por fspecii.
Todo el código UI original, arquitectura y diseño pertenecen a sus respectivos autores.
Este proyecto añade herramientas optimizadas para Windows y mejoras orientadas a producción musical.

ComponentAuthorLicenseLink
ACE-Step UI (base UI)fspeciiMITgithub.com/fspecii/ace-step-ui
ACE-Step 1.5 (AI model)ACE-Step TeamMITgithub.com/ace-step/ACE-Step-1.5
ProdIA-MAX (this fork)ElWalkiMIT
i18n system & translationsscruffynerfMITPR #1

🚀 What is ProdIA-MAX? / ¿Qué es ProdIA-MAX?

EN: ProdIA-MAX is a Windows-optimized fork of ACE-Step UI that bundles extra production tools: BPM/key detection, automatic lyric transcription, vocal/instrumental separation, LoRA training preparation, and one-click launchers — all wired to the ACE-Step 1.5 AI music generation engine.

ES: ProdIA-MAX es un fork de ACE-Step UI optimizado para Windows que incluye herramientas extra de producción: detección de BPM y tonalidad, transcripción automática de letras, separación vocal/instrumental, preparación para entrenamiento LoRA y lanzadores con un solo clic — todo integrado con el motor de generación musical IA ACE-Step 1.5.


✨ MAX Additions / Añadidos MAX

🎵 Generation & Audio

FeatureStatusDescription
Vocal Separation (Demucs)✅ BetaSeparate vocals/instrumental from any song
Vocal Reference TabDedicated vocal reference panel
Independent Audio StrengthsSeparate reference + source strength sliders
Audio Codes SystemFull semantic audio code pipeline — extract, apply & condition generation
Mic Recorder + Audio CodesRecord voice → auto-extract Audio Codes + optional Whisper transcription
Whisper Model SelectorChoose Whisper model (tiny→turbo) with download status indicators
Process + Whisper / Solo ProcesarTwo-button workflow: extract codes with or without lyric transcription
Chord Progression EditorInteractive chord builder with scale-aware suggestions and audio preview
1/4 Time SignatureAdded 1/4 time signature option to all time signature selectors

🤖 AI Assistant & UX

FeatureStatusDescription
Chat AssistantBuilt-in AI assistant (OpenRouter/local LLM) for music production guidance
LLM Provider SelectorSwitch between OpenRouter models or local LLM endpoints
Language SelectorFull i18n — switch UI language (ES/EN/+more)
Sidebar Info PanelClick info icons → scrollable info panel in sidebar (no clipping)
Professional Audio PlayerEnhanced player with playback speed, shuffle, repeat modes
Song Cards & Drag-DropVisual song cards with drag-and-drop to playlists
Resizable Chat PanelDrag to resize chat assistant panel
Style EditorEdit and manage generation style presets

🔧 Production Tools

FeatureStatusDescription
Audio Metadata TaggingID3 tags (title, artist, BPM, key) in MP3s
Edit MetadataEdit BPM, key, time signature, title in-app
LoRA Quick UnloadOne-click unload on collapsed LoRA panel
Time Signature LabelsProper 1/4, 2/4, 3/4, 4/4, 5/4, 6/8, 7/8, 8/8 notation
Prepare for Training✅ BetaQuick button to prep songs for LoRA training
VRAM SafetyGeneration locked during Demucs separation
BPM & Key Detectiondetectar_bpm_clave.py — batch detect BPM/key
Lyric Transcriptiontranscribir_letras.py — Whisper-based transcription
Caption ToolsApply/truncate training captions automatically
One-click Windows launchers.bat scripts for setup, launch, and cleanup

🏆 Best Quality Settings / Mejores Ajustes de Calidad

IMPORTANT / IMPORTANTE: Audio generation quality is dramatically better using the base model at 124 inference steps compared to turbo mode (8 steps). While turbo is faster (~3-5s per song), the base model at 124 steps produces significantly richer, more coherent, and higher-fidelity audio.

La calidad de generación de audio es MUCHÍSIMO mejor usando el modelo base a 124 pasos en comparación con el modo turbo (8 pasos). Aunque turbo es más rápido (~3-5s por canción), el modelo base a 124 pasos produce audio significativamente más rico, más coherente y de mayor fidelidad.

SettingTurbo (fast)Base (recommended)
Modelturbobase
Inference Steps8124
Speed (30s song)~3-5s~45-90s
QualityGoodExcellent
CoherenceDecentVery high
Recommended forQuick previewsFinal production

📋 Requirements / Requisitos

RequirementMinimumRecommended
OSWindows 10 / Linux / macOSWindows 11
GPU VRAM4 GB12 GB+
Python3.103.11
Node.js1820+
CUDA12.8
Disk Space~15 GB (required models)~40 GB (all models)

🔧 Installation / Instalación

What you MUST install manually / Lo que DEBES instalar manualmente

You only need 3 things installed before running ProdIA-MAX. Everything else is automatic.
Solo necesitas 3 cosas instaladas antes de ejecutar ProdIA-MAX. Todo lo demás es automático.

1. Python 3.11

EN: Download from python.org. During installation, check "Add Python to PATH" — this is critical.
ES: Descarga desde python.org. Durante la instalación, marca "Add Python to PATH" — esto es esencial.

⚠️ Python 3.11 is strongly recommended. The pyproject.toml locks to ==3.11.* and some dependencies (like flash-attn) only have pre-built wheels for 3.11.

2. Node.js 18+

EN: Download from nodejs.org (LTS recommended). This powers the backend server and both frontends.
ES: Descarga desde nodejs.org (recomendado LTS). Esto ejecuta el servidor backend y ambas interfaces.

3. NVIDIA GPU + CUDA Drivers

EN: You need an NVIDIA GPU with updated drivers. CUDA 12.8 compatible drivers are required — PyTorch will be installed with CUDA 12.8 support automatically. Without an NVIDIA GPU, the app cannot generate music.
ES: Necesitas una GPU NVIDIA con drivers actualizados. Se requieren drivers compatibles con CUDA 12.8 — PyTorch se instalará con soporte CUDA 12.8 automáticamente. Sin GPU NVIDIA, la app no puede generar música.

💡 Git is only needed if you clone the repo. If you download the ZIP, Git is not required.
💡 Git solo es necesario si clonas el repo. Si descargas el ZIP, no necesitas Git.

Verify installation / Verificar instalación

python --version    REM Should show 3.11.x
node --version       REM Should show v18+ or v20+
nvidia-smi           REM Should show your GPU and driver version

⚡ Quick Start / Inicio rápido

Windows

REM Option 1: Launch everything (Classic UI + Pro UI + AI engine)
iniciar_todo.bat

REM Option 2: Launch Pro UI only (recommended)
iniciar_pro.bat

Linux / macOS

chmod +x iniciar_todo.sh
./iniciar_todo.sh

What opens / Qué se abre

LauncherURLDescription
iniciar_todo.bathttp://localhost:3000Classic UI (+ Pro UI on :3002)
iniciar_pro.bathttp://localhost:3002Pro UI only (recommended)

🚀 First Run — What Happens Automatically / Primera ejecución — Qué pasa automáticamente

EN: The first time you run iniciar_todo.bat or iniciar_pro.bat, the script will automatically:

ES: La primera vez que ejecutes iniciar_todo.bat o iniciar_pro.bat, el script automáticamente:

StepWhat it doesTime (est.)
1️⃣Creates Python virtual environment (.venv inside ACE-Step-1.5_/)~10 sec
2️⃣Installs Python dependencies via pip install -r requirements.txt — PyTorch (CUDA 12.8), Transformers, Gradio, etc.~5-15 min
3️⃣Installs Node.js dependencies via npm install in 3 folders (ace-step-ui, ace-step-ui/server, ace-step-ui-pro)~2-5 min
4️⃣Downloads AI models from HuggingFace (~9.6 GB for required models)~10-30 min
5️⃣Starts all services and opens your browser~30 sec

⏱️ Total first run time: ~20-50 minutes depending on your internet speed. Subsequent runs start in ~30 seconds.
⏱️ Tiempo total primera ejecución: ~20-50 minutos dependiendo de tu velocidad de internet. Las siguientes ejecuciones arrancan en ~30 segundos.

A marker file .deps_installed is created after successful installation — subsequent runs skip the install step unless requirements.txt changes.


🧠 AI Models / Modelos de IA

All models are downloaded from HuggingFace automatically on first run. They are stored in ACE-Step-1.5_/checkpoints/.

Required models (auto-downloaded) / Modelos requeridos (descarga automática)

ModelSizeVRAMPurpose / Propósito
acestep-v15-turbo~4.5 GB~4 GBDefault generation model (diffusion transformer)
vae~500 MB~1 GBAudio encoder/decoder
Qwen3-Embedding-0.6B~1.2 GB~1 GBText encoder for lyrics/tags
acestep-5Hz-lm-1.7B~3.4 GB~4 GBLanguage model for music structure
Total~9.6 GB

Optional models (manual download) / Modelos opcionales (descarga manual)

Use verificar_modelos.bat to check which models you have and download extras:
Usa verificar_modelos.bat para ver qué modelos tienes y descargar extras:

ModelSizeVRAMPurpose / Propósito
acestep-v15-base~4.5 GB~4 GBBase model — 124 steps, highest quality ⭐
acestep-v15-sft~4.5 GB~4 GBFine-tuned variant
acestep-v15-turbo-shift1~4.5 GB~4 GBTurbo variant 1
acestep-v15-turbo-shift3~4.5 GB~4 GBTurbo variant 3
acestep-v15-turbo-continuous~4.5 GB~4 GBContinuous generation mode
acestep-5Hz-lm-0.6B~1.2 GB~3 GBSmall LM (less VRAM)
acestep-5Hz-lm-4B~7.8 GB~12 GBLarge LM (best quality lyrics)

💡 VRAM recommendation: 8 GB VRAM for turbo mode, 12 GB+ for base model at 124 steps with the 4B language model.
💡 Recomendación VRAM: 8 GB VRAM para modo turbo, 12 GB+ para modelo base a 124 pasos con el modelo de lenguaje 4B.

All models come from HuggingFace repos under ACE-Step/ — no account or token needed.
Todos los modelos vienen de repos HuggingFace bajo ACE-Step/ — no se necesita cuenta ni token.


📁 Project Structure / Estructura del proyecto

ProdIA-MAX/

├── ACE-Step-1.5_/              # ACE-Step 1.5 AI engine (original, MIT)
│   ├── acestep/                #   Core Python inference code
│   ├── requirements.txt        #   ★ Python dependencies (PyTorch, Gradio, etc.)
│   └── ...

├── ace-step-ui/                # Base UI fork from fspecii (original, MIT)
│   ├── server/                 #   Express + SQLite backend (Node.js)
│   ├── package.json            #   Node.js dependencies (backend + frontend)
│   └── ...

├── ace-step-ui-pro/            # ProdIA-MAX Pro UI (React 19 + Vite + Tailwind v4)
│   ├── src/                    #   TypeScript source code
│   │   ├── components/         #     React components (views, create, ui, layout)
│   │   ├── services/           #     API clients, chord service, etc.
│   │   ├── i18n.ts             #     Internationalization (EN/ES)
│   │   └── App.tsx             #     Main application entry
│   ├── package.json            #   Node.js dependencies (frontend)
│   └── ...

├── i18n/                       # i18n utilities & locale files
│   ├── locales/                #   Translation JSON files
│   └── utils.py                #   i18n helper scripts

├── IMAGES/                     # Screenshots & product images
│   └── main.png                #   Main product screenshot

├── SecurityScan/               # Security tools
│   ├── escanear_seguridad.bat  #   Run security scan (Windows)
│   ├── limpiar_kms.bat         #   KMS cleanup
│   └── LEEME.md                #   Security scan documentation

├── backups/                    # Session backups

│── ★ MAX Production Tools ──────────────────────────────
├── aplicar_captions_v3.py      # Auto-apply training captions
├── detectar_bpm_clave.py       # BPM & key detection (librosa)
├── transcribir_letras.py       # Lyric transcription (Whisper)
├── transcribir_letras_v2.py    # Lyric transcription v2
├── truncar_captions.py         # Caption truncation helper
├── check_genres.py             # Genre validator
├── check_tensors.py            # Tensor/model inspector

│── ★ Launchers (Windows) ───────────────────────────────
├── iniciar_todo.bat            # One-click launcher (backend + frontend + AI)
├── iniciar_pro.bat             # Launch Pro UI only
├── verificar_modelos.bat       # Model verification
├── limpiar_datos_usuario.bat   # Clean user data
├── desinstalar.bat             # Uninstall / cleanup
├── transcribir_letras.bat      # Transcribe lyrics launcher

│── ★ Launchers (Linux/macOS) ───────────────────────────
├── setup.sh                    # First-time setup
├── iniciar_todo.sh             # One-click launcher
├── iniciar_pro.sh              # Launch Pro UI only
├── iniciar_acestep_ui.sh       # Launch base UI
├── verificar_modelos.sh        # Model verification
├── limpiar_datos_usuario.sh    # Clean user data
├── desinstalar.sh              # Uninstall / cleanup
├── detectar_bpm_clave.sh       # BPM & key detection
├── transcribir_letras.sh       # Transcribe lyrics

│── ★ Documentation ─────────────────────────────────────
├── README.md                   # This file
├── SINGER_LIBRARY_SPEC.md      # Singer library specification
├── i18nHowTo.md                # i18n guide (English)
└── i18nHowTo_es.md             # i18n guide (Spanish)

🎼 Audio Codes System / Sistema de Códigos de Audio

ACE-Step uses semantic audio codes — tokens at 5Hz that encode melody, rhythm, structure, and timbre holistically. Each code (<|audio_code_N|>, N=0–63999) represents 200ms of audio.

How to use:

  1. Load audio in Source Audio → click "Convert to Codes" → codes appear in the Audio Codes field
  2. Or use the Voice Recorder → "Process + Whisper" extracts codes automatically
  3. Write a different text prompt (e.g., change genre/style)
  4. Adjust audio_cover_strength (0.3–0.5 = new style with original structure)
  5. Generate → the model follows the code structure but with your new style
PropertyValue
Rate5 codes/second (200ms each)
Codebook size64,000 (FSQ: [8,8,8,5,5,5])
30s song~150 codes
60s song~300 codes
QuantizerFinite Scalar Quantization (1 quantizer, 6D)

Note: Individual codes encode ALL musical properties simultaneously (melody + rhythm + timbre). You cannot isolate rhythm from melody at the code level, but you can use codes as structural guidance while the text prompt controls style/instrumentation.


🎙️ Voice Recorder Workflow / Flujo del Grabador de Voz

  1. Record your voice/humming/beatbox
  2. Choose a Whisper model (tiny→turbo) — download status shown with ✓/↓ indicators
  3. Click "Procesar + Whisper" → extracts Audio Codes AND transcribes lyrics
  4. Or click "Solo Procesar" → extracts Audio Codes only (faster, no transcription)
  5. Click "Aplicar" → codes + lyrics are applied to the Create Panel
  6. The generation will use your voice structure as a musical guide

�️ Scripts Reference / Referencia de Scripts

Main Launchers / Lanzadores principales

ScriptWhat it does / Qué hace
iniciar_todo.batFull launch: installs deps if needed → starts AI engine (port 8001) → backend (port 3001) → Classic UI (port 3000) → Pro UI (port 3002) → opens browser
iniciar_pro.batPro launch: same as above but skips Classic UI — only AI engine + backend + Pro UI
iniciar_todo.shLinux/macOS equivalent of iniciar_todo.bat

Utility Scripts / Scripts de utilidad

ScriptWhat it does / Qué hace
verificar_modelos.batChecks which AI models are installed and offers to download missing ones from HuggingFace
transcribir_letras.batLaunches the lyrics transcription tool (uses Demucs for stem separation + Whisper for transcription). 6 quality modes available
limpiar_datos_usuario.batDeletes user data only (database + generated audio). Keeps models, code, and dependencies
desinstalar.batFull uninstall — removes .venv, all node_modules, database, and generated audio. Keeps models and source code

Internal Scripts (ACE-Step-1.5_/) / Scripts internos

ScriptWhat it does / Qué hace
_start_gradio_api.batStarts only the Gradio AI engine on port 8001
_start_backend.batStarts only the Node.js backend on port 3001
_start_frontend.batStarts only the Classic UI on port 3000
iniciar_acestep.batInteractive menu: standalone Gradio UI (port 7860), API mode, model download, advanced config

Production Tools / Herramientas de producción

ScriptWhat it does / Qué hace
detectar_bpm_clave.pyBatch detect BPM and musical key of audio files
transcribir_letras.pyTranscribe lyrics from audio using Whisper
aplicar_captions_v3.pyApply training captions to audio files for LoRA preparation
truncar_captions.pyTruncate captions to fit training requirements
check_genres.pyValidate genre tags in audio files
check_tensors.pyInspect model tensor files

🔌 Ports / Puertos

PortServiceDescription
8001Gradio APIACE-Step AI music generation engine
3001Express.js BackendAuth, model management, audio serving, database
3000Classic FrontendLegacy React UI (Vite dev server)
3002Pro FrontendProdIA-MAX Pro UI (Vite + React 19 + Tailwind v4)
7860Gradio Native UIOnly used with iniciar_acestep.bat standalone mode

All ports are killed automatically before launch to avoid conflicts. The scripts handle this for you.
Todos los puertos se liberan automáticamente antes del lanzamiento para evitar conflictos.


❓ Troubleshooting / Solución de problemas

"Python not found" / "No se encuentra Python"

  • Make sure Python 3.11 is installed and added to PATH
  • Run python --version in a terminal to verify
  • If you have multiple Python versions, ensure 3.11 is the default

"Node.js not found" / "No se encuentra Node.js"

  • Install Node.js from nodejs.org
  • Restart your terminal after installing
  • Run node --version to verify

"CUDA out of memory" / "Sin memoria CUDA"

  • Use a smaller language model (0.6B instead of 1.7B or 4B)
  • Use turbo mode instead of base model
  • Close other GPU-intensive applications
  • Check your VRAM with nvidia-smi

Models not downloading / Los modelos no se descargan

  • Check your internet connection
  • Run verificar_modelos.bat to manually download models
  • Models are downloaded from HuggingFace — no VPN restrictions needed
  • If download was interrupted, delete the incomplete folder in ACE-Step-1.5_/checkpoints/ and retry

Port already in use / Puerto en uso

  • The launcher scripts auto-kill processes on required ports
  • If you still get errors, manually close any process using ports 8001, 3001, 3000, or 3002
  • On Windows: netstat -ano | findstr :PORT_NUMBER then taskkill /PID <PID> /F

Audio doesn't generate / No se genera audio

  • Ensure the Gradio API started successfully (check the terminal window titled "Gradio API")
  • Wait for the message "Running on local URL: http://127.0.0.1:8001" before trying to generate
  • The first generation takes longer as PyTorch compiles kernels

Fresh start / Empezar de cero

REM Remove all installed dependencies (keeps models and code)
desinstalar.bat

REM Then relaunch — everything reinstalls automatically
iniciar_todo.bat

�📄 License / Licencia

This project is distributed under the MIT License.
Este proyecto se distribuye bajo la Licencia MIT.

  • The original ACE-Step UI code remains © fspecii under MIT.
  • The ACE-Step 1.5 model and backend remain © ACE-Step Team under MIT.
  • New additions and modifications in this fork are © ElWalki under MIT.

See ACE-Step-1.5_/LICENSE for the full license text.


Built on top of the amazing work of fspecii and the ACE-Step team.
i18n system & translations contributed by scruffynerf.
Stop paying subscriptions. Start creating with ACE-Step.