Models with an embedded LoRA are a workaround for achieving the effects provided by a LoRA with a Core ML based Stable Diffusion pipeline. An alternative to the direct use of LoRAs is necessary because Apple's Core ML Stable Diffusion implementation does not provide native support for LoRAs.

The LoRA effect in these models works best when the "trigger-words" that were used in the LoRA's training are included in the text prompt. Conversely, the effect of the embedded Lora can be minimized by including the trigger-words in the negative prompt, but in general, there will be residual effects from the LoRA present in any image generated by these models.

A merge is most practical when the model's basic characteristics match the LoRA's intended effect. A LoRA trained on a particular anime character should be mated with an anime-focused base model, a NSFW niche LoRA should be embedded in an NSFW model, etc.

The listings below indicate the LoRA and the base model used in each merge. The folder for each merged model typically contains builds at several image sizes, sample image file(s), and a text file with the suggested trigger-words from the embedded LoRA. (Some LoRAs don't use trigger-words.)

If there is a merge you'd like me to run for you, leave a note in the Community area here. I'm open to anything that is SD-1.5 type. Let me know the LoRA, the base model, and what size(s) you'd most like.

SD-2.1 type base models are not, and will not be, included here. They requie pipeline modifications on my end, and they are not always suppored by the end use applications.

The merges here were made with Automatic1111 and the resulting .safetensors models were tested there. The models were then converted to Core ML and tested with Mochi Diffusion They should also work in a standard Swift CLI pipeline that uses a recent build of ml-stable-diffusion as a base.

These models should work with Imaga2Imange, but they do not include the ControlledUnet.mlmodelc component required for use with ControlNets.

Take care in opening the sample image and trigger-word files here as they may contain words and images that you find offensive. Caveat emptor.