apple
ControlNet directory too large ...
The ml-stable-diffusion library now supports ControlNet, which is an excellent addition thanks to the contributors. However, I have noticed that when specifying multiple types using --convert-controlnet, the file size in the ControlNet directory becomes too large (as shown in the attached image). Is there a solution ?
That appears to be the sizes they need to be. I have tried converting starting from both fp32 and fp16 .bin files and they both produce the same 723.1 result. To pile on further, it appears that you must also convert the models for each output size you want to work with -- a Canny.mlmodelc at 512x512, another Canny.mlmodelc at 768x768, etc. And then there is also original vs split_einsum. FWIW, these models look to be the same size (or double at fp32) for a python diffusers pipeline, but you may not need multiple image size versions there.
That appears to be the sizes they need to be. I have tried converting starting from both fp32 and fp16 .bin files and they both produce the same 723.1 result. To pile on further, it appears that you must also convert the models for each output size you want to work with -- a Canny.mlmodelc at 512x512, another Canny.mlmodelc at 768x768, etc. And then there is also original vs split_einsum. FWIW, these models look to be the same size (or double at fp32) for a python diffusers pipeline, but you may not need multiple image size versions there.
🙂