🌸 ZhuangFangyiAE Character LoRA (SDXL / Illustrious)

A character LoRA trained on Zhuang Fangyi from Arknights Endfield ✨

This new version was fully trained manually by me with love and a lot of experimentation ❤️
While there was an older version trained through the Civitai, this release is my own self-trained version/

Maybe it doesn't fully surpass the older version yet 🌿

The model was trained on carefully selected in-game screenshots with cleaned captions and optimized settings for SDXL anime checkpoints.

✨ Main Features

• Preserves the original in-game appearance

• Stable face and hairstyle generation

• Consistent horns, tail and outfit details

• Semi-realistic anime game-render style

• Works very well with cinematic lighting prompts 🎬

The dataset includes:

• close-up face shots;

• upper body shots;

• full body screenshots;

• multiple camera angles;

• different poses and lighting conditions.

🧩 Recommended Models

Works best with:

• Illustrious XL

• JANKUTrainedNoobaiRouwei

• NoobAI XL

• Pony-based SDXL checkpoints

• Semi-realistic anime SDXL models

🔑 Trigger Word

ZhuangFangyiAE

⚙️ Recommended Settings

Sampler: dpmpp_2m
Scheduler: karras
Steps: 28-40
CFG: 5.5-6.5

🎛️ Recommended LoRA Strength

Model strength: 0.65-0.9
Clip strength: 0.4-1.0

🌸 Example Character Prompt

ZhuangFangyiAE, (1girl:1.3), long hair, multicolored hair, green hair, black hair, dragon horns, dragon tail, pointy ears, elegant dress, bare shoulders, black gloves, covered breasts, covered body, (full outfit:1.3), green clothes, white clothes, yellow eyes, detailed eyes, green hakama pants, wide pants

InGame scene

game screenshot, in-game render, 3d anime render, high quality game graphics, physically based rendering, global illumination, volumetric lighting, screen space reflections, ambient occlusion, cinematic lighting, sharp focus, detailed face, sharp eyes, high detail textures, subsurface scattering, dynamic lighting 

📝 Notes

• Lower LoRA strengths usually provide cleaner and sharper results.

• Epochs around 5-7 provided the best visual quality during testing, The 6th epoch was taken.

• Works especially well with cinematic prompts and game-render style prompts.

• Euler-based samplers may produce softer or blurrier results compared to DPM++ samplers.

🛠️ Training Information

• Base architecture: SDXL

• Training type: LoRA

• Rank: 32

• Resolution: 1024

• Dataset: in-game screenshots

• Captions: manually cleaned WD14 captions

• Optimized for anime / game-render checkpoints

Thank you for checking out this project 💖