Abstract
We define a broader family of corruption processes that generalizespreviously known diffusion models. To reverse these general diffusions, wepropose a new objective called Soft Score Matching that provably learns thescore function for any linear corruption process and yields state of the artresults for CelebA. Soft Score Matching incorporates the degradation process inthe network and trains the model to predict a clean image that after corruptionmatches the diffused observation. We show that our objective learns thegradient of the likelihood under suitable regularity conditions for the familyof corruption processes. We further develop a principled way to select thecorruption levels for general diffusion processes and a novel sampling methodthat we call Momentum Sampler. We evaluate our framework with the corruptionbeing Gaussian Blur and low magnitude additive noise. Our method achievesstate-of-the-art FID score $1.85$ on CelebA-64, outperforming all previouslinear diffusion models. We also show significant computational benefitscompared to vanilla denoising diffusion.