Inference time-step selection

[Turlan]

I noticed that the timesteps are [atan(σ_max, 1.3, 1.0, 0.6)] for inference. How were these specific numbers selected? Are there any guiding principles behind them?

[zhengkw18]

These are empirical choices (e.g., the σ_max parameter inherits from sCM). We did not carefully tune them.