CUBIGCorp/Gyeongbokgung_Palace_Models_Based_Synthetic_Data

This repository provides four LoRA (Low-Rank Adaptation) adapters designed to be used with
Stable Diffusion XL (SDXL) for generating images of selected cultural heritage assets in Gyeongbokgung Palace.

The collection focuses on front-view, full-structure representations, preserving architectural identity and visual consistency.

⚠️ These LoRAs are not standalone models and must be used with an SDXL base model.

---

Included Heritage Assets

Designation	Asset Name (KR)	Asset Name (EN)	ID	Description
국보	경회루	Gyeonghoeru Pavilion	224	Two-story pavilion built over a pond, used for royal banquets, state examinations, and ceremonial events.
보물	경복궁 풍기대	Wind Streamer Pedestal of Gyeongbokgung Palace	847	Pedestal used to observe wind direction and strength, reflecting Joseon-era meteorological practices.
보물	향원정	Hyangwonjeong Pavilion	1761	Hexagonal two-story pavilion on an island in the rear garden pond, used for rest and appreciation.
보물	아미산 굴뚝	Chimneys in Amisan Garden	811	Decorated royal chimneys connected to Gyotaejeon’s ondol system, featuring symbolic relief motifs.

---

Model Overview

• Model Type: LoRA adapters for image generation
• Base Model: Stable Diffusion XL (SDXL)
• Scope: Selected Gyeongbokgung Palace heritage assets (4)
• Viewpoint: Front view, full architectural composition
• Data Type: Synthetic image–based fine-tuning

Each LoRA is trained independently per asset, ensuring high structural fidelity and reduced cross-asset interference.

---

How to Use (Diffusers)

import torch
from diffusers import StableDiffusionXLPipeline

base_model = "stabilityai/stable-diffusion-xl-base-1.0"
lora_path = "Gyeonghoeru_Pavilion.safetensors"  # change per asset

pipe = StableDiffusionXLPipeline.from_pretrained(
    base_model,
    torch_dtype=torch.float16,
).to("cuda")

pipe.load_lora_weights(lora_path)
# pipe.fuse_lora()  # optional

prompt = "a photo of Gyeonghoeru Pavilion, front view, architectural structure"
image = pipe(prompt, num_inference_steps=30).images[0]
image.save("output.png")

---

Training Data Description

Data Source

• Dataset
  • (Korean) Synthetic Data-based Source Image Augmentation Dataset for Cultural Heritage Images in Gyeongbokgung Palace
  • (English) Cultural Heritage Images of Gyeongbokgung Palace: Synthetic Data-based Source Image Augmentation Dataset
• Reference Source: National Heritage Portal of Korea

Data Characteristics

• Metadata referenced from publicly available cultural heritage records
• Synthetic images generated internally as part of a government-funded AI digital heritage research project
• No indiscriminate web-crawled data was used

Data Scope

• Viewpoint: Front view
• Composition: Full architectural structure
• Language: Korean metadata
• Typical Scale (per asset):
  • Original reference images: ~4
  • Synthetic images: ~96

---

Personal Data & Sensitive Information

All training data was designed to exclude personal and sensitive information.

• Potentially sensitive elements (e.g., people, license plates, camera equipment) were detected and removed or anonymized during preprocessing.
• The dataset contains no personally identifiable information (PII).

---

Intended Use

Primary Objectives

• Improve cultural heritage–aware image generation quality
• Preserve architectural structure and visual identity of selected assets
• Support public-sector and cultural heritage–related AI applications

Intended Applications

• Cultural heritage content creation (text-to-image)
• Tourism and educational materials
• Reference-level restoration visualization and simulation
• Public administration, archival, exhibition, and education
• AI research and technical validation

---

Out of Scope / Usage Restrictions

This model collection is not suitable for:

• Structural or engineering analysis
• Architectural design or restoration planning
• Legal, technical, or historical authentication

These LoRAs do not replace:

• Measured drawings
• Scholarly research
• Official restoration or conservation judgments

All outputs are reference-level visual materials only.

---

LoRA-Specific Notes

• The pretrained data composition of the SDXL base model is not modified.
• Each LoRA is trained for:
  • A single heritage asset
  • A fixed dominant viewpoint (front view)

Model behavior and limitations depend directly on the training scope and synthetic data strategy.

---

Limitations & Risks

• Excessively long or highly stylized prompts may introduce:
  • Structural distortion
  • Alteration of fine architectural details

Careful prompt design is recommended to maintain fidelity.

---

Technical Summary

• Base Model: stabilityai/stable-diffusion-xl-base-1.0
• Method: LoRA (selected UNet and Text Encoder modules)
• Resolution: 1024 × 1024
• Training: Mixed precision, parameter-efficient fine-tuning
• Artifacts: LoRA weights (.safetensors)