0.1.6

• Audio-to-Audio Variation Mode — The Foundation-1 Generate node now accepts an optional audio input. Connect any ComfyUI AUDIO output (LoadAudio, previous generation, etc.) to create variations/interpretations guided by your prompt. Controlled by a new init_noise_level parameter (0.01–1.0): lower values stay close to the input, higher values produce more creative interpretations.
  Bug Fixes
• Fixed unload_after_generate CPU resume — The model was incorrectly resuming to CPU instead of CUDA on the second generation after offloading, producing silence. The device field now correctly preserves the original target device.
• Fixed "Free Memory" button not working — ComfyUI's native "Free Memory" button now correctly fully unloads the model from RAM when unload_after_generate=True.
• Improved memory cleanup — GPU memory cleanup now runs double garbage collection passes (handles objects whose destructors reference other collectible objects) and explicitly nulls out model sub-components (pretransform, conditioner, model backbone) before deletion.
• Fixed misleading No module named 'stable_audio_tools' error — Both the loader and generator nodes now provide clear install instructions (pip install stable-audio-tools --no-deps) when the package is missing, instead of a bare import error.
  Changed
• offload_to_cpu() resets torch.compile state — Compiled model wrappers are cleared on offload to avoid stale cached graphs when resuming.