- Ship the latest VAE checkpoint and PCA basis refit on its latents
(sliders: Click, Punch, Bright, Sub, PC5).
- Security: torch.load(weights_only=True), np.load(allow_pickle=False)
— pca_params.npz now stores pc_names as a plain string array so it
loads without pickling — and add per-endpoint rate limiting
(generate 10/min, spectrogram 30/min).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Auto-detects GPU at startup, loads BigVGAN from HuggingFace Hub
- Falls back to Griffin-LIM on CPU
- Loads fine-tuned BigVGAN weights from models/vocoder/ if present
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
