ml-stable-diffusion icon indicating copy to clipboard operation
ml-stable-diffusion copied to clipboard

Can one coreml package integrate multiple resolutions?

Open czkoko opened this issue 2 years ago • 0 comments

I noticed that the coreml model cannot switch resolutions arbitrarily like ckpt. As a result, I need to repeatedly generate multiple huge model packages.

I use Guernika Model Converter to convert ckpt model developed by @GuiyeC, It is very easy and works with his Guernika.

My computer is M1, so I give priority to supporting ANE's SPLIT_EINSUM format, which requires 2.14G storage space, but SPLIT_EINSUM only supports 512x512.

So I need to convert 512x768 and 768x512 in ORIGINAL format.

The three basic resolutions require 6G+ storage space, and ControlNet also requires multiple models with different formats and resolutions.

In this way, it not only takes up space, but also switches between different models, and it is troublesome to find the corresponding ControlNet.

Is it possible to use multiple resolutions with only a single coreml model like ckpt in the future?

czkoko avatar Apr 21 '23 14:04 czkoko