Riffusion + ControlNet trained on 150 songs in Slakh2100 dataset for 3 days on NVIDIA A10. Generate audio from spectrogram controlled by denoised Sobel spectral edgemap, with text conditioning