Difference between split-einsum models?
coreml-anything-v3.0 / split-einsum:
- anything-v3.0-current_split-einsum_first-VAED.zip
- anything-v3.0-first_split-einsum_current-VAEE.zip
- anything-v3_split-einsum.zip
Why are there three of these models instead of one? Why "VAEE" and "VAED"? Is the last one current/current, first/first, or something else? Which one should I use?
I believe that VAED and VAEE have different VAEs burned in from the third one, which is the one you should download as it is the default version.
Last year anything-v3.0 was the first model released. Later, anything-v3-better-vae was released. But now anything-v3-better-vae is published as anything-v3.0.
The CoreML model converted from the currently published anything-v3.0 has a problem that the color of the generated image using Neural Engine becomes very strange.
anything-v3.0-current_split-einsum_first-VAED.zip avoids this problem by using the first model VAEDecoder.
anything-v3.0-first_split-einsum_current-VAEE.zip is the CoreML model converted from the first model with the current model VAEEncoder.
anything-v3.0-current_split-einsum_first-VAED.zip is my recommended model.