Mitigating Exposure Bias in Discriminator Guided Diffusion Models

Abstract

Diffusion Models have demonstrated remarkable performance in image generation. However, their demanding computational requirements for training have prompted ongoing efforts to enhance the quality of generated images through modifications in the sampling process. A recent approach, known as Discriminator Guidance, seeks to bridge the gap between the model score and the data score by incorporating an auxiliary term, derived from a discriminator network. We show that despite significantly improving sample quality, this technique has not resolved the persistent issue of Exposure Bias. Exposure bias refers to the discrepancy between the input data during training and inference phases and leads to diminished sample quality in diffusion models. We propose SEDM-G++, which incorporates a modified sampling approach, combining Discriminator Guidance and Epsilon Scaling. Our proposed framework outperforms the current state-of-the-art in unconditional image generation.