Trained on 44 images I generated using https://civitai.com/models/2217715?modelVersionId=2485111 with a specific prefix prompt and processed with a detailer. Trained for 114 epochs for Z-Image, 50 for Anima.

Klein 4B/9B

The LoRA works well both in editing and text-to-image mode. The 4b version can be pushed a little stronger while the 9b version is already quite strong at 1.0. It can give different variations depending on prompt, I like to trigger it with

"@mikkoani illustration with muted colors and complex gritty linework". Adding "anime" also works very well.

Anima

Can be used with base model with CFG=4, Steps 16-30 or using https://civitai.com/models/2466415/cosmos-predict25-2b-base-distilled-extracted-dmd2-lora with CFG=1 and Steps=8

Works best with er_sde, euler, res_multistep and heun.

The example images have almost always "muted colors" and "complex linework" in the prompt and usually no quality modifiers.

Z-Image

Works with both Z-Image-Base + https://civitai.com/models/2362961?modelVersionId=2741469 or with Z-Image Turbo.

The LoRA doesn't have a fixed trigger word (was trained with @mikkoani but seems to be irrelevant) but can produce different styles depending on the prompt.

My favorite triggers are:

"mature illustration with muted color and complex linework"

"fashion sketch"

"mature anime illustration"

you can make the images creepier with "horror themes"

With Z-Image-Base you get more freedom and less reliance on the prompt. With Z-Image-Turbo you'll need to be a bit more convincing with your prompt to get the desired effect (unless you desire semi-realism, that comes out of the box)

Best used at 1.0, but can be pushed higher, especially with ZIT.