Common Diffusion Noise Schedules and Sample Steps are Flawed

• DOI: 10.48550/arxiv.2305.08891
• Publication Date: 2023-01-01
• Authors: Shanchuan Lin; Bingchen Liu; Jiashi Li; Xiao Yang
• Keywords: Implementation; Gaussian Noise

We discover that common diffusion noise schedules do not enforce the last timestep to have zero signal-to-noise ratio (SNR), and some implementations of samplers start from timestep. Such designs are flawed reflect fact model is given pure Gaussian at inference, creating a discrepancy between training inference. show design causes real problems in existing implementations. In Stable Diffusion, it severely limits only generate images with medium brightness prevents generating very bright dark samples. propose few simple fixes: (1) rescale schedule terminal SNR; (2) train v prediction; (3) change sampler always timestep; (4) classifier-free guidance prevent over-exposure. These changes ensure process congruent inference allow samples more faithful original data distribution.