lumicero/yae-miko-waifu-diffusion-lora-db

Waifu Diffusion – Frieren Dreambooth LoRA

Dreambooth LoRA adapter for hakurei/waifu-diffusion to generate Yae Miko character images.

_{lmyyaemiko, 1girl, pink_hair, fox_ears, japanese_clothes, cinematic, masterpiece, best quality}

Quickstart (Diffusers)

import torch
from diffusers import StableDiffusionPipeline

base_id = "hakurei/waifu-diffusion"
lora_id = "lumicero/yae-miko-waifu-diffusion-lora-db"

pipe = StableDiffusionPipeline.from_pretrained(base_id, torch_dtype=torch.float16).to("cuda")
pipe.load_lora_weights(lora_id)

prompt = "lmyyaemiko, 1girl, pink_hair, fox_ears, japanese_clothes, cinematic, masterpiece, best quality"
image = pipe(prompt, num_inference_steps=28, guidance_scale=7.0,
             cross_attention_kwargs={"scale": 0.8}).images[0]
image.save("out.png")

Usage tips

• Trigger word: lmyyaemiko, 1girl, pink_hair, fox_ears
• Recommended LoRA scale: 0.6–0.9
• Steps: 20–35
• Instance prompt: lmyyaemiko

Training details

• Dataset: CyberHarem/yae_miko_genshin (IMG+TXT stage3)
• Resolution: 512
• LoRA rank: 16
• Steps: 2500
• LR: 2e-4

Evaluation Score

The model was evaluated using the evaluation tool sdeval from deepghs.

25 images generated from the base model and lora were compared with 250 ref images.

Model	CCIP (Mean ± Std)	BikiniPlus (Mean ± Std)	AICorrupt (Mean ± Std)
Base	0.03440 ± 0.088784	0.764154 ± 0.123648	0.868899 ± 0.270314
LoRA	0.99808 ± 0.004636	0.750238 ± 0.102336	0.960905 ± 0.069958