Practical effects of full precision vs. half precision weights on Stable Diffusion

Hello everyone,

I am aware that the Stable Diffusion model for use with the Diffusers library is available both in full precision and half precision (fp16) versions, and that using the half precision version saves VRAM.

However, are there any other practical differences? Is the half precision model slower, or do the generated images tend to have less quality? When would you want to choose the full precision model over the half precision one?

Thanks!

1 Like