While working on another lora I noticed that Pony doesn't really know the horizontal splits position (it does know vertical splits however).<h2 id="versions-a45nk221v">Versions</h2><ul><li>1.0: LyCORIS/Dora (requires Automatic1111 1.9+) with extended dataset and improved tagging</li><li>0.6: Standard Lora with smaller dataset</li></ul>Adetailer is recommended for both versions<h2 id="tags-llw5ucven">Tags</h2>The model knows the following 3 main positions (bold are the tags)<ol><li>horizontal splits</li><li>vertical splits</li><li>horizontal splits,upside-down</li></ol>With the following supportive tags you can generate most splits positions:<ul><li>top-down bottom-up</li><li>arched back</li><li>upside-down</li><li>on stomach</li><li>on back</li><li>bent over</li><li>...</li></ul>Perspective tags<ul><li>front view</li><li>from behind</li><li>from below</li><li>from above</li><li>from side</li><li>...</li></ul>NegativesIf you notice that the model goes too much 3d, put "3d" in the negative prompt. Some of the images in the dataset leaned a bit too much into 3d for my liking, but I wanted the model to learn the concept better."watermark" helps agains the watermark and logos, but unfortunately the model still struggles with it a bit (kinda burnt myself out manually editing the dataset images to get rid of background and watermarks, but as it turned out, this also worsened the model T_T)<s>I've been told that the model quite frequently seems to create two females instead of one. If that is the case for you, specifying </s><s>1girl</s><s> and/or </s><s>solo female</s><s> in positive and </s><s>2girls</s><s> in the negative prompt might help. Most likely this happens because there is one image in the training dataset, which has two females in a vertical pose, because I wanted the model to learn to also support multiple persons.</s><s>I will try to fix it in the next version</s><h3 id="resolutions-iogbm51bd">Resolutions</h3>I mostly use the following resolutions, not sure if others work:<ul><li>1216x832</li><li>1344x768</li><li>1024x1024 (sometimes the model tries to fit in the full body making the legs unproportionally short)</li><li>768x1344</li><li>832x1216</li></ul><s>I used a new training method and I'm not entirely content with the results yet, so I will probably retrain the model using my old method, but it will take some time to adjust the dataset etc.</s><s>For now enjoy the model, it should work pretty well for most cases</s>Happy prompting, I'm looking forward to your creations!

Splits positions [Pony XL]

-Changed from standard Lora to new Dora type since it seems to perform better-Changed trigger word from "upside-down splits" to standard "horizontal splits" combined with the "upside-down" tag-model should have less of an 3D effect-better prompt adherance

Files with SHA256: e6758430dab348803c25732373a77aeabdcbf65f8d17b76a228b63fdf8c405b5