fp32

[viyjy]

Hi, thanks for this great work.
I have a question regarding encode_latents_in_fp16. If we set encode_latents_in_fp16=False to use fp32, do we expect lower performance compared to fp16? I have tried both and found that images generated by fp32-based model have lower quality than fp16-based model. Is that expected? Thanks

Best

[Landanjs]

Hi, how much worse is the quality? Do you have examples?

We haven't checked this ourselves, but it is possible that the VAE quality at fp32 could be worse if the VAE was trained on fp16.