llustrious:

• Strength - 0.8 - 1.2;

• 896x1152;

• Upscaling - hires. fix x1.6 d.0.55 latent(nearest-exact) + face adeteiler;

• Suggested prompts: traditional media.

• I recommend using Manticore (second preview) or something similar to vpred.

Z-Image:

• Strength - 0.8 - 1;

• Upscaling - SeedVR2;

• Suggested prompts: Cinematic anime art, semi-realistic rendering, painterly shading, soft light gradients, dramatic rim lighting, smooth materials, volumetric atmosphere, depth of field, high polish finish, modern fantasy style, detailed character art, filmic color grading.

Stabilization:
I see that some users are having difficulty obtaining stable anime image generations on Z-image and ZIT. Here are some simple stabilization and basic preprocessing techniques.

ZIT and Z-image do not fully comprehend the Illustrious anime we are accustomed to. While working with Z-image, I noticed that focusing on clear ZIT tags for clothing and surroundings significantly stabilizes the overall result. For example, prompt with [Clothes: Black matte bra, black shorts. Straps on the thighs, straps on the stomach, collar] will give a more stable result than without these tags.

Similarly, with the environment, a prompt with environment tags, for example: [Background: Black matte background, indoors bedroom] will give a more accurate result than without these tags.
All these tips are not universal and, often, may not work on other lora or give the wrong result.

Using FlowMatch? Here are my experimental settings:

• shift = 1.5

• steps = 12

• base_shift = 0.4

Similarly, for Color Grading:

• Temperature: +12

• Contrast: +15

• Gamma: 0.9