Time-Mender: An AI Cultural Relic Image Restoration Platform Based on SAM and LaMa

Time-Mender is an open-source web platform for cultural relic and art restoration. It integrates the Segment Anything Model (SAM), LaMa image restoration, and Stable Diffusion technologies, built on FastAPI, to provide an end-to-end solution for the digital preservation of cultural heritage. This platform lowers the technical barrier for digital restoration of cultural relics and supports restoration workers and researchers in efficiently carrying out digital restoration tasks.

Digital preservation of cultural relics and artworks is an important issue in contemporary cultural heritage inheritance. Traditional manual restoration is costly and requires professional skills, while breakthroughs in AI technology in image restoration have opened up new possibilities. The Time-Mender project emerged as a response, integrating cutting-edge AI technologies into a unified web platform to provide a complete digital restoration toolchain for cultural relic restoration workers and researchers.

Technical Architecture

Time-Mender is built on the FastAPI framework with a modular design. The backend uses Python deep learning libraries, and the frontend adopts a lightweight HTML/CSS/JS solution, balancing maintainability and scalability.

Core Engines

1. SAM Segmentation Model: Meta AI's general-purpose segmentation model. It generates high-quality segmentation masks via user prompts (point selection/box selection) to accurately locate restoration areas. Its zero-shot learning capability adapts to various types of cultural relic damage.
2. LaMa Image Restoration: A high-performance restoration model from Samsung Research. Based on a Fourier convolution architecture, it excels at filling large missing areas and producing visually coherent restoration results.
3. Stable Diffusion Enhancement: An optional feature that supports creative reconstruction of severely damaged areas, detail supplementation, and multi-solution generation. Professional evaluation is required to ensure historical authenticity.

Time-Mender provides a complete user workflow:

• User Management: Registration and login, with independent storage space and isolated data.
• Image Upload and Preprocessing: Supports single/batch uploads, with automatic format handling and size adjustment.
• Interactive Restoration Interface: Visual operations to select restoration areas, adjust parameters, and preview results.
• Multi-Engine Selection: Choose between LaMa or Stable Diffusion restoration engines based on needs.
• Result Export: Supports single/batch download of restoration results.

Deployment Methods

• Local Quick Start: Clone the repository → Create a virtual environment → Install PyTorch → Install dependencies → Download SAM weights → Start the service (uvicorn).
• Production Deployment: It is recommended to use uvicorn/gunicorn as the ASGI server, combined with nginx reverse proxy and systemd daemon to ensure stability.

Data Security

• Data Separation: Runtime data (uploaded images, restoration results) is stored in data/, and model weights are stored in models/ for easy backup and migration.
• Privacy Protection: It is recommended not to commit data/ and models/ to version control to avoid repository bloat and data leakage.

Application Prospects

• Museums/Archives: Assist in digital collection restoration and improve large-scale processing efficiency.
• Academic Research: The open-source architecture facilitates secondary development and algorithm verification.
• Education: An intuitive demonstration platform to help understand the principles of AI restoration.

Limitations

• The frontend interface is basic, and user experience needs optimization.
• Stable Diffusion integration requires additional models and configurations.
• Performance for ultra-large-scale image processing needs further optimization.

By integrating cutting-edge models such as SAM and LaMa into an easy-to-use web platform, Time-Mender lowers the technical barrier for digital restoration of cultural relics and enables more institutions and individuals to participate in cultural heritage protection. With project iterations and community contributions, it is expected to become an important open-source tool in the field of cultural relic restoration.