ArtRestore-Diffusion: A Unified Style-Consistent Generative AI System for Artwork Restoration

ArtRestore-Diffusion is a generative AI architecture for restoring damaged areas of historical paintings. It achieves style-consistent restoration through mask-adaptive FiLM conditioning, artist style embedding injection, and retrieval-augmented denoising. It also provides an interpretable block-wise attribution mechanism and a style certificate for quantifying style consistency. The system aims to address the problems of time-consuming and costly traditional restoration, as well as the lack of targeting and style consistency in existing AI methods.

Historical artworks are prone to damages such as cracks, scratches, large-area missing parts, and color fading. Traditional restoration relies on expert manual labor, which is time-consuming and costly. Existing AI restoration methods have two major limitations: 1. They handle all damage types uniformly, lacking targeting; 2. They cannot guarantee artist style consistency, leading to违和 repair areas. ArtRestore-Diffusion addresses these challenges, aiming to create a unified restoration architecture that can handle multiple damage types while maintaining style consistency.

ArtRestore-Diffusion adopts a latent diffusion architecture, including four core innovative components:

1. Mask-Adaptive FiLM Conditioning: Handles binary (cracks, scratches, etc.) and continuous (color fading) masks through a single CNN+FiLM module, enabling unified processing of multiple damage types;

2. Artist Style Embedding Injection: Uses AdaLN and cross-attention to inject style embeddings, ensuring consistent restoration style;

3. Retrieval-Augmented Denoising: Retrieves similar samples from a style sample memory bank to assist denoising;

4. Block-Wise Attribution Mechanism: Generates attribution maps to trace generated areas back to source samples, improving interpretability.

The style certificate is an innovation of the project, providing formal boundaries for deviations from the artist's style:

• Calculates the distance between the style embedding of the restored area and the centroid of the artist's style manifold;

• Sets deviation thresholds to quantify style consistency;

• Provides violation rate statistics for quality assessment, offering a quantifiable standard for digital restoration.

Testing Validation: Includes 13 smoke tests, 20 unit tests, 19 domain benchmark tests, 11 ablation experiments to evaluate component contributions, and performance analysis tools.

Current Status: In the pre-experiment infrastructure phase; unit tests and benchmark tests have passed, while performance metrics (LPIPS, DISTS, style certificate) need to be verified after training.

Future Directions: Collect large-scale datasets, train the complete model, conduct expert validation, and develop user-friendly tools.

Technical Contributions:

1. Uniformly handles multiple damage types, simplifying deployment;

2. Block-wise attribution mechanism improves interpretability;

3. Style certificate establishes a quantifiable quality standard.

Application Scenarios: Digital restoration in museums, preview for artwork market evaluation, art history research, educational communication, etc., contributing to cultural heritage protection.