The LOCON Version of My CattleyaV2 LORA. Same dataset so get them both and compare. :P I can't simply overstate how therapeutic deleting Rana from the dataset was. It was a zen activity I would recommend for the whole family.
This works fine with: <lyco:CattleyaV3:0.7>, cattleya_qb, low_ponytail
The low ponytail part is not necessary but seems to stabilize her hairstyle, a couple of images with her hair loose might have slipped in without me properly tagging them.
I was only able to isolate one outfit but decided to split her neck protector as she uses it with some other outfits. The outfit triggers are:
cattleya_purple_highleg_leotard: Her standard combat leotard outfit.
cattleya_gorget: Her neck guard/protector. This one can be a bit iffy but it is stable enough, didn't know how that thing was called(gorget) had to dig in danbooru. The tagger didn't detect it either.
As always, I will add the tag summary with the dataset so people can experiment further.
Description
FAQ
Comments (4)
hey dude you work is amazing i am noob in training lora and i have trained 10-20 loras but out of which 8-9 are only decent enough ..well can you please guide me how should i train consistent lora that work fine on high weights...
thanks in advance
You can check the guide i made. https://civitai.com/articles/138/making-a-lora-is-like-baking-a-cake Otherwise the most important things are properly cleaning the dataset to remove extra elements, watermarks and other nastiness and I always use 100+ images dataset which makes the process a bit more forgiving if any of my images are crap.
Saw the sample images of your LORAs. They look nice but I know that game so I imagine you are suffering from scarcity in the dataset. In that case I would recommend to make a synthetic dataset. Create a Lora regardless of if it ends up crappy and use it to img2img the first generation dataset to generate extra images for a second generation LORA. Just make sure to pay extra attention to discard images with artifacts or bad hands in the generated dataset.
@knxo thank you so much dude 😇