Disentangling Total-Variance and Signal-to-Noise-Ratio Improves Diffusion Models
The authors introduced a Total-Variance/Signal-to-Noise-Ratio (TV/SNR) framework to independently control these two parameters in diffusion models, overcoming limitations of existing schedules. They found that using a constant TV schedule consistently improves performance when the variance grows exponentially, and they proposed an Exponential Inverse Sigmoid SNR (ISSNR) schedule for state-of-the-art sampling efficiency.