This my first Style LoRA trained with the creations of "Double Deck". It's also provide some characters from his work, Even if this can be a pretty niche style, I'm hoping you may find what you where searching for.
Even with a very small dataset, it do work with realistic and furry style.
The keyword is 'double deck, monster girl' but you can use specific character like :
- 'mushroom girl, plant girl', 'anthro girl, frog girl', 'anthro girl, raccoon girl' for animal/plant characters
- 'robot girl', 'android', 'tesla girl' and 'assaultron' to generate robot
- 'silent hill' plus 'pyramid head' or 'bubble head nurse', which are more chaotic during generation
- 'white and black' for the original style, or 'colored picture' to force painted variant of the style
Some arguments may exist but aren't usefull enough or may break the generation, like 'multiple breasts' or others...
For the Strength of this LoRA, it seem to work OK on 1, but It work best by lowering a bit values if there is artefact during generation.
Description
This is a first variant of the DoubleDeck Lora trained on Pony Diffusion V6 XL, this is an experimental version for feedback purpose. Actions made to increase training quality :
- Balanced training data to have equal representation
- Cleaned pictures from any parasite character
FAQ
Comments (9)
Wow, new fetish obtained.
adding another to the collection like the general grievous of gooning
Looks pretty nice mang, but any way to drop the 800mb to the usual 200? I don't make LoRAs so no idea really
2nd this - I'll stick with the 1.5 model for now. That 800mb one is just asking too much from my computer.
Hey, it's my first SDXL LoRA, I didn't really think of it when I try to train it, I previously saw other LoRA for SDXL going up to 800MB so I through it was normal. I will check some documentation about that and see if it's possible to compress it.
@Lazius Will thank you mang, it really seems like an interesting style, but my 2tb look smaller each day
There you go https://pixeldrain.com/u/jhy8bz5m
Resized with sv_ratio at 15, Average Frobenius norm retention: 83.00% | std: 0.089
Comparison: https://i.imgur.com/9m60uCO.jpeg (Top row are the originals)
@Lazius You can train styles with dim 8 and less with XL, resulting in pretty small files, there's no quality loss, most of the large file size from loras is random numbers from the network initialization, and making high dim loras, especially for XL, mostly increases that amount of random initialization values the file will contain (though it is better to make high dim and trim later than initially train it at low dim, but 99% of what we train doesn't require high dim to achieve good results)
iirc this one is dim 1 https://civitai.com/models/307792/calm-ponyxl-size14mb?modelVersionId=356015
Dim 8 https://civitai.com/models/375389/kumagorow-style
The rest of my loras are all dim 16, which is still kinda wasteful, I myself find it hard to resist just cracking up the numbers but its not needed
I have been waiting for this Lora for so long, thanks my good man. May nothing but good things come your way!