Trigger: xlomrblng

This LoRA produces a painterly effect based on paper marbling patterns. Works best for landscapes since the training images are landscapes. Only rarely converts people, but should manage their clothes & accessories.

I’m hoping to add people to the training set and do a new version.

It’s a synthetic style: no artist names or works were used in training.

Only tested in automatic1111 webui.

Generally I’ve been using the trigger near the start of the prompts, with the LoRA weighted at default 1.0, although 0.83 <---> 1.21 seem to be usable too.

Negative prompt generally:

toy,replica,scale model,nude,naked

because of some figurine type images I didn’t want, and to be SFW on CivitAI.

Sampler: worked well with

DPM++ 2M Karras around 40 steps

DPM++ 3M SDE Exponential around 25 steps

CFG scale: 7 seems to work well

Mostly used Hires.fix at either 1.25 or 1.5 to increase the detail a bit. No other adjusters or post in use. Hires.fix steps 25<--->30, Latent (nearest-exact), denoising 0.7.

Model:

mostly Crystal Clear XL

https://civarchive.com/models/122822/crystal-clear-xl

although I did get some interesting results with

CinEro XL | Photomatic Cinegraphy - Moody Creative Cinematic

https://civarchive.com/models/151568?modelVersionId=321398

For example prompts you should be able to pick them off the showcase images.

You can use my TIs at

https://civarchive.com/models/154898/marblingtixl

with the LoRA, although it does make the paper marbling rather intense :-)

It works, though, so I imagine other TIs would too.

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

Some rambling...

This LoRA is partly an experiment to see how different (or not) a style LoRA and a style TI embedding are if trained on the same dataset with kohya_ss using similar settings.

My first attempt at this sort of thing was a while back when I couldn’t get any bleeping thing to train at all. I’d just about given up in frustration when konyconi encouraged me to try making a TI from his BohoAI dataset. Eventually managed it using automatic1111 webui training.

konyconi’s original LoRA for Boho style (for SD v1.5):

https://civarchive.com/models/51966/bohoai-konyconi

The Boho training data can be found at konyconi’s tutorial here:

https://civarchive.com/articles/771/tutorial-konyconi-style-lora-update

And my TI version of the Boho style is here (for SD v1.5):

https://civarchive.com/models/61239/bohoti

Since then I’ve switched to kohya_ss for training and done a few TIs here on Civitai. I’ve been using the idea of things made from paper marbling patterns

https://en.wikipedia.org/wiki/Paper_marbling

to test various training options, and this LoRA is the latest output.

It uses the same image dataset as my MarblingTIXL, from where you can download the trainset + TI kohya_ss settings if you’re interested:

https://civarchive.com/models/154898/marblingtixl

Today’s publication is a LoRA made from those same images. Rather different output, although you can see the paper marbling with both the TI and the LoRA.

The TI gives the, umm, sense of the patterns in a sort of idealised way. Whereas the LoRA here uses the elements from the training images in a much more direct way. I think you can probably get away with less images for TI training (provided they’re good/careful ones of course).

However, I can see I’ll need to increase the image count for this type of LoRA to provide more variability and more instances of particular things I want to see in the LoRA output. e.g. lighthouses - only got one training image with a lighthouse, so the LoRA wants to show that particular style in preference to all other types of lighthouses. And people - I’d like to be able to apply the effect to fantasy portraits.