Commit Graph
5 Commits
Author SHA1 Message Date
diodiogod e6b9a5ab80 Version 4.8.27
Centralize numba compatibility system with smart testing

Technical implementation:
- Replace 15+ duplicate JIT workaround blocks with unified system
- Add utils/compatibility/numba_compat.py with startup testing
- Implement fallback for environments without compatibility module
- Test librosa.stft() at startup and apply workarounds only when needed
- Eliminate contradictory startup messages between init and nodes
- Maintain backward compatibility with legacy Python 3.13 fixes
2025-09-09 00:40:06 -03:00
diodiogod 1e61db1fca Version 4.8.20
Fix Python 3.13 RVC and vocal separation compatibility

Technical fixes:
- Fix RVC f0 method dependencies (pyworld, torchfcpe) for complete functionality
- Fix vocal separation numba/librosa compatibility issues in Python 3.13
- Fix MDX23C model path resolution and Karafan engine integration
- Fix model substitution and fallback error handling
- Improve default model selection (MelBand over slow ONNX on CPU)
- Add comprehensive safe wrapper functions for librosa operations
2025-09-06 22:55:31 -03:00
diodiogod cd853b70eb Fix Python 3.13 librosa/numba compatibility in vocal separation
- Add safe_stft(), safe_istft(), and safe_resample() fallback functions
- Replace problematic librosa calls in UVR spec_utils.py with safe versions
- Try librosa first, fall back to torchaudio on numba compilation errors
- Use librosa-compatible defaults for torchaudio (Hann window, center=True)
- Fixes 'get_call_template' AttributeError in Python 3.13
- Preserves audio quality while maintaining compatibility
2025-09-06 20:55:55 -03:00
diodiogod 42f4fa3acd Fix MDX mono/stereo dimension mismatch in vocal separation
- Move mono-to-stereo conversion to beginning of initialize_mix()
- Ensures both is_ckpt and else branches handle mono audio correctly
- Prevents concatenation errors with mixed channel dimensions
2025-09-06 18:20:05 -03:00
diodiogod aa6c679449 Implement SCNet SOTA architecture for vocal separation (10.08 SDR)
MAJOR NEW FEATURE:
- Full SCNet implementation from ZFTurbo's Music-Source-Separation-Training
- SCNet achieves 10.08 SDR - significantly better than previous models
- Paper: https://arxiv.org/abs/2401.13276.pdf

TECHNICAL IMPLEMENTATION:
- Added complete SCNet architecture with Sparse Compression Network
- Dual-Path RNN with frequency/time processing
- Multi-band sparse down/up-sampling layers
- Transformer-based separation network
- Compatible interface with existing UVR5/RVC framework

NEW SOTA MODELS ADDED:
- model_scnet_xl_ihf_sdr_10.08.ckpt (BEST OVERALL - 10.08 SDR)
- model_scnet_masked_xl_ihf_sdr_9.82.ckpt (Noise reduction variant)
- model_vocals_mdx23c_sdr_10.17.ckpt (Best vocals only)
- denoise_mel_band_roformer_sdr_27.99.ckpt (Best denoising - 27.99 SDR\!)

ARCHITECTURE FEATURES:
- Sparse compression for efficient processing
- Multi-band frequency processing (low/mid/high)
- Advanced dual-path RNN with bidirectional LSTM
- Feature conversion with FFT/IFFT operations
- Encoder-decoder with skip connections and fusion layers

INTEGRATION:
- Automatic SCNet model detection by filename
- Proper parameter handling (no aggressiveness for SCNet)
- GPU acceleration support
- Compatible caching system
- Updated model tooltips with new SOTA recommendations

LICENSE: MIT (ZFTurbo/Roman Solovyev) - Free to use and modify
2025-08-08 12:01:22 -03:00