Fixing model settings
#4
by Yntec - opened
If your model converted to diffusers doesn't work on the Inference API, or it produces bad quality outputs, plugging it on the Model Converter and making a "Fp32" "No Ema" version will be enough to fix it. However, for a model that is +7.7GB in size, like this one, you can't do that because you will run out of memory before it finishes saving.
What I did instead was using the model OldFish 1.1 that is only 1.8GB, to create an Alpha 1.0 merge in supermerger, and used the settings below to create a no ema version of this model that the inference API can use to produce images of good quality:
