Do not use these styles with other SDXL models, you just won't see any effect, These styles are for Pony Models only!
It's kind of an experiment, like why not? Why not train the model on just one image?
Styles are as stable as possible for just one image
Styles are a bit unstable but it's not related to LoRA, it's more related to the Pony Diffusion V6 XL model.
Recommended LoRA Strength: 1 (Strength can be reduced to 0.7 below this value I do not recommend it)
About OIS-s
JPEG compression artifacts were removed and training settings and network size were changed, LoRA is now half the size and style quality is also better (Styles are more similar to the original images)
Styles are based on images created by the AI
Description
Repeats: 100
Epoch: 20
Image Count: 1
Network Rank: 32
Network Alpha: 32
Batch size: 1
Clip skip: 1
Training steps: 2000
Resolution: 1024x1024
Train Resolution: 1024x1024
Model: ponyDiffusionV6XL_v6
WD14 Captioning: SmilingWolf/wd-v1-4-moat-tagger-v2
LoRA Weight: [1]
FAQ
Comments (6)
The images were taken from Twitter/X user @patata1216
Style 1 - https://twitter.com/patata1216/status/1759352091339108443
Style 2 - https://twitter.com/patata1216/status/1759259946213110075
Style 3 - https://twitter.com/patata1216/status/1758997614320840771
Style 4 - https://twitter.com/patata1216/status/1759943832995529010
Style 5 - https://twitter.com/patata1216/status/1759872302190342642
Style 6 - https://twitter.com/patata1216/status/1760234384190578829
Fascinating experiment. I guess Pony just has so much data that it really latched onto something familiar.
So, because it was a single image, did it train really fast? Could you share your settings?
@Jellai Not quite that fast, 1 image repeated 100 times is the same as repeating 100 images 1 time ='(
The first LoRA I created was trained on 8 images (I didn't have more anyway).
You need at least 4 images to create a style (Stable than with 1-2 images)
I have a theory or something, when Style LoRA is combined with a model it changes almost nothing. it doesn't change the concept or anatomy, I trained styles on images with only male or female but it didn't stop the model from creating both genders or even Pony/Furries.
Style LoRA is kind of like a Filter, nothing more.
Сharacter LoRA can already change the concept or anatomy and a little bit of style.
My LoRAs are more in control of the style than trying to replicate it
I also substituted that there are words that make a big difference in style realistic, lineart, pixel art, traditional media, monochrome, greyscale
100_foldername
1 x100 x20 = 2000
Image Count: 1
Batch size: 1
Repeats: 100
Epoch: 20
Training steps: 2000
Using less than 2000 steps I wouldn't recommend it.
I can share a link to Kohya_ss - Configuration file if you need?
nice style
Looks interesting, especially style 2
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.