TL;DR a style Lora to generate sharp looking cyber ladies (as a result of my adventure in using Stable Diffusion)

Works with DPM++ 2M or Euler a (both with Karras), can be activated with "cyb3rlad1es", stack not too bad on PonyXL or SDXL with standard resolutions (1024x1024 or my usual 832x1216).

Can be used at strength 1 if alone, but 0.8 is enough most of the time

First (and probably only) attempt at training a Lora.

The long version of the story (links to lot of good resources for beginners)

To try and better understand all the small stuff about Stable Diffusion, i first installed Automatic1111 and started with the classical AnythingXL model to try and generate cool women in cyberpunk settings (because why not).

Very fast, i found both Style Selector and GhostXL to trigger more cyberpunk in there (with Neon Punk style), together with Adetailer (for face, eyes and lips).

But since i liked some results from pure AnythingXL, i used the provided Checkpoint merger to do "AnyGhostXL" (0.5/0.5 merge of previously mentioned model).

At this point, it was high time to start using Lora and two of them found their way in all my tries (because i saw them in pictures i liked):

• Detail Tweaker XL (at 1.5 strength)

• Doctor Diffusion's Negative XL (at -1 strength)

Since they became present all the time, it was worth baking them in my own AnyGhostXL. Using the SuperMerger, both found their way in with these default strengths, creating "AnyGhostLora".

Happy with this SDXL checkpoint, which i used together with a custom VAE (SharpSpectrum), i started looking for help on prompt and finally found PonyXL...

Merging PonyXL with my own model was not giving good results, that's why i started using PonyXL as a refiner instead, switching at 0.7, and after a while, instead of using the base PonyXL, i switched to DucHaiten-Pony. And of course, i started mixing some ExpressiveH in order to get nice expressions.

But it still didn't feel good enough, that's when Kuro CAT minimalistic style got into play to really get the style I imagined.

Together with Textual Inversions (negativeXL and unaestheticXL), it was time to generate "en masse".

Using Dynamic Prompts, i started getting a good lot of cool ladies and decided in was the right moment to switch out of my custom checkpoint AnyGhostLora (based on SDXL) to PonyXL based model for generation (and avoid pushing this merged stuff here to post pictures).

I have included the wildcards I used and here is the prompt (without the PonyXL stuff).

1girl, looking at viewer, __body__, __viewpoint__, __subject__,  __eyes__, (__hair__), __cut__, {1-2$$__modifiers__}, {1-2$$__accessory__}, (__legwear__), __expression__, (__lips__:1.1), Fisheye lens, (Bokeh) <lora:Minimalistic_Style_Pony:0.8> BREAK __location__, cyberpunk style, __lighting__

After generating a good close up and with ControlNet together with OpenPose, i generated a bunch of ladies for Lora training.

Using Kohya, and PonyXL as base model, you have the result here :D

To tag the dataset (~80 pictures), i could have extracted using pillow the prompts used, but to get a more generic bunch of tags, i preferred using WD14.

That's all folks, thanks for reading!

NB: At one point, maybe i'll post AnyGhostLora too since it still work well with this lora ^^;

AnyGhostLora is on Civitai too now :D