CLAUDE.md

DiffRhythm Project Guide

Commands

• Run web interface: python app.py
• Install dependencies: pip install -r requirements.txt

Code Style Guidelines

• Imports: Standard lib first, third-party second, local imports last
• Type Annotations: Use tensor dimension annotation style
  • float["b n d"] where b=batch, n=sequence length, d=dimension
• Naming: Functions/variables: snake_case, Classes: PascalCase, Constants: UPPER_CASE
• Utils: Use existing helper functions (exists(), default())
• Error Handling: Use try/except with specific error messages

Project Structure

• app.py - Main entry point with Gradio web interface
• diffrhythm/ - Core library components:
  • config/ - Model configs in JSON/INI format
  • g2p/ - Grapheme-to-phoneme for multiple languages
  • infer/ - Inference code for music generation
  • model/ - DiT and CFM model architecture

Architecture Notes

• Models take lyrics with timestamps (LRC) as input
• Uses audio/text style prompts to guide music generation
• Employs diffusion models that align with lyrics
• Maintain dimension consistency when working with tensors