Takashi Shibuya

Spectral Prior for Reducing Exposure Bias in Diffusion Models

Yuya Kobayashi, Masato Ishii, Yuhta Takida, Takashi Shibuya, Yuki Mitsufuji

Diffusion models typically suffer from error accumulation during iterative sampling, commonly referred to as exposure bias. We reveal systematic frequency-dependent discrepancies between training and inference, which can be interpreted as frequency-dependent SNR error. Crucially,…Read more