Monte Carlo Score Matching for Image Generation

Score-based models are state-of-the-art generative models for image generation. We propose a novel loss namely the Monte Carlo Score Matching (MCSM) loss as an approximation of the original score matching loss. MCSM leverages a Taylor-series expansion of the score function to approximate the expensive calculation involved in computing the trace of the Jacobian of the score function. MCSM is competitive with models trained using the Sliced-Score Matching (SSM) loss. We validate the efficacy of the proposed technique in terms of negative log-likelihood and Fréchet Inception distance (FID) on MNIST and CelebA datasets, respectively. In particular, we show that FID of images generated with models trained using MCSM loss is on par with, and in some cases, better than, sliced score-matching for image generation.