README_SPACE.md

metadata

title: HeartMuLa Music Generation
emoji: 🎵
colorFrom: purple
colorTo: pink
sdk: docker
sdk_version: latest
app_file: app.py
pinned: true
duplicated_from: HeartMuLa/HeartMuLa-oss

HeartMuLa Music Generation Space

Generate music from lyrics and style tags using the HeartMuLa family of open-source music foundation models.

Features

• Music Generation: Convert lyrics + style tags → audio via two-stage pipeline
• HeartMuLa LLM: Frame-by-frame audio token generation with style control
• HeartCodec: High-fidelity flow-matching codec (48 kHz output)
• Multiple Model Sizes: 3B, 7B, and 300M versions available

Setup

The Space will automatically download and set up the required models on first run.

Usage

1. Enter your lyrics in the text field
2. Add style tags (e.g., "pop, upbeat, energetic")
3. Adjust generation parameters:
   • Duration: Length of generated music (5-60 seconds)
   • Temperature: Creativity level (0.1-2.0)
   • CFG Scale: Style control strength (1.0-3.0)
   • Top-K: Sampling parameter (10-100)
4. Click Generate Music to create your track

Model Information

• HeartMuLa-RL-oss-3B-20260123: RL-tuned version with improved style control (recommended)
• HeartCodec-oss-20260123: Optimized audio decoding quality

Performance

• RTF ≈ 1.0 (real-time inference speed)
• 48 kHz sample rate output
• Supports multiple languages

References

• Paper
• GitHub
• Discord