Question

[ShiLiangfan11]

When using diffusion for image super-resolution, having too many inference steps can cause the details to deviate more from the real image.

[cswry]

Each step of the diffusion process randomly samples noise, which is the fundamental cause of the resulting randomness. Naturally, the more inference steps, the greater the randomness. If you want to reduce randomness in multi-step inference, you can refer to our recently work CCSR.