Score-of-Mixture Training: Training One-Step Generative Models Made Simple via Score Estimation of Mixture Distributions
This work addresses the problem of efficient and effective training of generative models for computer vision tasks, particularly for those who require simple and stable training methods.
The authors tackled the problem of training one-step generative models and achieved competitive or superior results to existing methods, with experiments on CIFAR-10 and ImageNet 64x64. Their approach, Score-of-Mixture Training, showed stable training and required minimal hyperparameter tuning.
We propose Score-of-Mixture Training (SMT), a novel framework for training one-step generative models by minimizing a class of divergences called the $α$-skew Jensen--Shannon divergence. At its core, SMT estimates the score of mixture distributions between real and fake samples across multiple noise levels. Similar to consistency models, our approach supports both training from scratch (SMT) and distillation using a pretrained diffusion model, which we call Score-of-Mixture Distillation (SMD). It is simple to implement, requires minimal hyperparameter tuning, and ensures stable training. Experiments on CIFAR-10 and ImageNet 64x64 show that SMT/SMD are competitive with and can even outperform existing methods.