Real&3D&Anime is multi-style base models, not limited to realism alone.

I don't want to destroy the anime concept of Anima. So i finetune this model in anime style by default.I'm aimming to use high-quality real cosplay and scenery photo data to achieve realistic style, and improve anime concept's anatomical\scene composition, through its realistic concept.

Recommended settings：

• postive prompts

  masterpiece,best quality,newest,high resolution,

• postive prompts for realistic

  photo \(medium\) ,[ photo \(medium\) : 0.5–3]

• negative prompts

• worst quality,low quality,lowres,signature,text,logo,european,mature female,lowres,e621,fat,saegyo sal,puffy cheeks,thick lips,

• Suggest parameters

  Sampling Method：ER SDE / Eulr / Heun / DPM++ 2M SDE

  Schedule Type：beta / simple
  cfg : 3~5（I often use 4.5 for realistic）
  shift：3
  

• Artist Style

  @ danbooru artist name

Part of Specially Trained Tags

Background: blurry foreground, blurry background (trained to help the model distinguish blurred areas)

Art Style: photo (medium), 3D, impasto, draft, flat color, no lineart, ···

Lighting: dark, chiaroscuro, night, dark background, backlighting, overexposure, ··· (Chiaroscuro delivers effects similar to camera flash; alternative tags are used in training to prevent the model from generating cameras)

Composition: partially underwater shot, partially underground shot, partially submerged, dutch angle, pov crotch, female pov, foreshortening, ···

Age: ···, mature female, old man, old woman, height difference, ···

Ethnicity & Creature: European, mermaid, monster, ···

Visual Effects: bloom, vignetting, motion blur, motion line, chromatic aberration, ···

Special Features: aegyo sal, puffy cheeks, entry plug (Aegyo sal refers to a popular eye makeup style widely seen in cosplay photos)

The following tags are not recommended for use:

1. solo: Removed from training datasets; its use results in 2D outputs. Replace with solo focus instead.

2. Duplicate standalone tags such as "shirt, red shirt": Synonymous duplicate tags have already been filtered out to improve prompt description accuracy.

Training Tool

I use (https://github.com/wochenlong/lora-scripts-next) for Full-finetuning. The tool is made by Chenkin (且听风吟，https://civarchive.com/user/Chenkin).And thanks for his technical help.