This is a model merge checkpoint based on PonyXL with some of my LORAs merged into. The focus is trying to preserve PonyXL's default set of styles and characters while also adding the ability to do photorealistic images as well as hopefully better monsters,vehicles, mechs, and landscapes. This model is mostly a test/learning expierience in model merging, but hopefully it will also be able to produce decent images. The current version is still very much in beta so expect some wonky results. In order to improve realism, I also merged in SXZ Luma and Pony Realism, in order to strengthen the ability to generate base pony art styles those models are either incapable of or have difficulty generating. I recommend checking each version's description as some may need different keywords and parameters. This model obviously inherits PonyDiffusion's license, as well as those of <a target="_blank" rel="ugc" href="https://civarchive.com/models/25831?modelVersionId=725032">SXZ Luma</a> and <a target="_blank" rel="ugc" href="https://civarchive.com/models/372465?modelVersionId=914390">Pony Realism</a>, and I'm unsure whether or not I need permission to post it to civitai to begin with. If you want to merge this model with another your perfectly fine to, though I would wait for a more stable version if you intend to do so.

Ponyspanded

Merged in the SXZ Luma and Version 2.2 of Pony Realism checkpoints to improve the model's ability to generate realistic images. Also merged in a few loras of cartoon and anime style artists to complement the Lora's of my own that I merged in. It seems to accept Luma and Pony Realism keywords as now too, which can be useful for negatives. To generate realistic images, I recommend using the DPM 2A sampler with a CFG around 2.5 and around 20 steps, and you should use "photo,realistic,cinematic" in your positive and "2d,cartoon,anime" in your negative. I found that embeddings can also help, as seen in the example images. To generate 2d images, put "photo,realistic,cinematic" and whatever style you want in the positive (I dont recommend any specific sampler of CFG for 2d images, it works well enough with Euler A anyway). It seems that it tends toward a cartoonish style with bright shading and thick outlines when doing 2d gens, so you may need to put "cartoon" or "cel shading" in the negatives. Most default pony styles seem to work, though for the 2d ones you will need to use the aforementioned negatives. In addition, this version seems to require stronger LORA weights than prior versions, you may need to pump some over 1.0 to get them to work properly.

Files with SHA256: 9d545988ce310cfcb482d0215dd50382e840dd833df0831a653a36c05d8d2486