See the 'about this version' on side about changes to the model.

This model merge is an attempt to improve the realism in pony models without sacrificing the notable versatility of the original model. The plasticky look of skin in some pony derived models is caused by a lack of detail that came from the original model as it wasn't trained on photos but illustrations with varying degrees of stylization. A novel method that I found is adding Perlin noise in the generation process to produce more detail in the image. So I created a lora out of Perlin noise and merge it with the Pony Realism model. After that merge, I did another one with the RealVisXL fp32 model. As a result the model is quite large at around 12gb so it can be pruned. (Pruned version uploaded)

The infamous scoring system associated with pony models are not mandatory but is still useful, so you can prompt the model more normally although the danbooru tag style is still preferred. The recommended sampler to use is DPM++ 2s A. If you want to improve the realism use source_photo in the prompt. Pony loras should work normally as I tested a couple of them but the output will be realistic in nature especially if the lora's strength is not too high. Also, it can do SFW images relatively easily by including the rating_safe tag in the positive prompt and rating_explicit in the negative.

The faces are often okay up to about 4 people, then they become more wonky when there is more. Enabling the restore face function should help in that situation. This model is still a work in progress. No embeddings or loras were used in the preview images.