KIMCAHLLIE/shinkai-style-lora

Shinkai-style LoRA (Stable Diffusion 1.5)

This repository contains LoRA weights fine-tuned on top of Stable Diffusion 1.5 to reproduce background-oriented anime-style visuals inspired by Makoto Shinkai.

---

🔑 Trigger Token

Use the following trigger token in your prompt:

shinkai_makoto_style

---

📊 Dataset

• Number of images: 25
• Resolution: 512x512
• Captioning: auto-generated captions + trigger token
• Format: imagefolder + metadata.jsonl

Dataset is included in this repository:

dataset/

---

🧠 Training Details

• Base model: runwayml/stable-diffusion-v1-5
• Method: LoRA fine-tuning (Diffusers)
• Training steps: 1500
• Batch size: 1
• Gradient accumulation: 4
• Mixed precision: bf16
• Environment: Google Colab (single GPU)
• Checkpoint resume used due to Colab runtime limits

---

🖼️ Inference Example (Diffusers)

import torch
from diffusers import StableDiffusionPipeline

base_model = "runwayml/stable-diffusion-v1-5"
lora_repo = "KIMCHALLIE/shinkai-style-lora"

pipe = StableDiffusionPipeline.from_pretrained(
    base_model,
    torch_dtype=torch.float16,
    safety_checker=None
).to("cuda")

pipe.load_lora_weights(lora_repo)

prompt = (
    "shinkai_makoto_style, rainy city street at dusk, "
    "cinematic lighting, detailed background, anime illustration"
)

image = pipe(
    prompt=prompt,
    num_inference_steps=45,
    guidance_scale=6.5
).images[0]

image.save("result.png")