SplitSplash Style LoRA
Bring your characters into beautifully illustrated aquatic worlds.
SplitSplash is a style LoRA trained on a curated collection of anime-inspired water illustrations, designed to create vibrant lakes, pools, oceans, and underwater scenes with clean painterly rendering and expressive lighting.
Recommended Prompting
For over-under shots, add terms such as:
split waterline composition
over-under composition
camera at the water surface
half above water, half below water
half-submerged
lower body visible beneath clear water
Description
One thing I noticed while testing is that using fewer trigger words seemed to produce more consistent results. This could simply be a quirk of Anima that I'm not familiar with yet, but when I included words like "water", the model often added bubbles around the character's head—even when that part of the image wasn't supposed to be underwater.
For the most consistent results, I found it worked best to use only the splitsplash trigger word and let the rest of the prompt describe the scene naturally.
FAQ
Comments (3)
"Half in water, half out of water" = partially submerged
1 trigger for the captions "splitsplash" is perfect. But all other words in your dataset should (must) be existing booru tags. Illustrious is booru style tagging: only. Your phrases are english yeah, but for the model you're training on some of those words are just: noise.
If I take your sample image as an example, if it was in a dataset, it should be captioned: splishsplah, wading, partially underwater shot, partially submerged
This is understandable by Illustrious.
It's just FYI ;) Thanks for the training.
N.B: The submerged tag you advise actually exists, but it means completely underwater. You see the confusion for Illustrious ;) Captioning is the best way to learn how to prompt ;)
Cheers. Good luck for next attempts.
Technically nothing stops you from using nonexistent tags or sentences on SDXL as long as that tag/sentence is actually represented well enough in your dataset, as the model needs to learn that tag from scratch (you already do this for trigger words after all). Problem is that there's no reason to do this if a tag already exists or comes close enough to what you want (so you wouldn't want to do something like "very long green hair" over just "long hair, green hair). If you only have a couple images then it'll overfit and turn out poorly regardless. In general nothing is stopping you from having multiple triggers, it's already what you do in multi-concept/style/character Loras anyways. In this case you are right that partially submerged should've been used since it's already well defined, but there is no real restriction on how you caption as long as you have enough images to generalize and not overfit.
The bigger issue imo is that CLIP on SDXL leaks tags a ton (ex. prompting for Diamond from Houseki no Kuni will often show a diamond gemstone, prompting for fruit hair ornaments might show the fruit itself etc), so using something like NL for concepts is a kinda debatable over just using multiple different single word triggers.
Thank you, that explains why i get better results when i say ( legs Submerged) then separately: legs, submerged,. and why when i just say SplitSplash, it has better results.