About the fourier time embedding type

[WHan-alter]

Dear Develops,

I noticed that in https://github.com/huggingface/diffusers/blob/17ecf72d4472f5dc11e7c86841c95898f6edbc0b/src/diffusers/models/unet_2d.py#L307-#L309, you scaled the samples with the timestep when the time embedding type is fourier. But you did not apply such normalization for the unet2d_condition and unet1d model. What are the differences between the scaled and the original sample? Yang Song's ScoreSDE has similar functions (they used the marginal prob std to normalize the model output). But I still need to get your idea regarding this.

Best, Wenkai