- Fix CFG 1.0 crash: auto-clamp to 1.1 (batch_size mismatch in attention layers)
- Extend CFG range from 5.0 to 10.0 for experimental genres
- Add comprehensive tag guide to README (genre, voice, mood, structure tags)
- Add CFG scale recommendations (2.0 pop, 3.0 rock, 4.0+ electronic)
- Document multi-language lyrics support
- Add VRAM/stability info from extensive testing
- Add .gitignore
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add fade-out post-processing (default 3s smooth fade)
- Dynamic model folder scanning (auto-detect all models)
- Smart quantization detection (reads config.json)
- Better defaults: CFG 2.0, duration 120s, max 240s
- Fix default tags to match official HeartMuLa format
- Reorder UI: model + quantize at top, cfg_scale prominent
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
All-in-one music generation + lyrics transcription for 16 GB GPUs.
Includes 4-bit/8-bit quantization, OOM fix, transformers 5.x compat,
torchtune RoPE fix, and soundfile replacement for torchaudio.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
