Efficient Unet architecture

[Bikesuffer]

Hi there,

I encountered some difficulties while attempting to remove certain blocks mentioned in the paper. Specifically, I'm having trouble removing one Resnet block in the Up-1 stage. Out of the three Resnet blocks present, I tried removing each one individually, but none of them seemed to work as expected.

In addition to the aforementioned Resnet block, I have successfully removed all the other blocks mentioned in the paper, based on the Stable Diffusion v1.5 model. However, I noticed that the number of parameters is currently 731M, which is lower than the value stated in the paper (848M).

Could you please provide more details about the efficient Unet architecture? This would greatly help me in gaining a better understanding of how to modify the architecture effectively.

Thank you.

[XiaoluDing]

In fact, I am a little skeptical that the laion data can be trained well using the diffusers. Hope the author could give you some convincing replies.