TL;DR
This paper proposes a high-fidelity framework for transforming delicate Korean embroidered artifacts into "living" digital assets. By combining optical scanning with Physically-Based Rendering (PBR) and structured metadata, the researchers created interactive 3D textures that preserve not just the image, but the complex materiality—luster, stitch depth, and translucency—of traditional textile heritage.
What Problem Does This Paper Address?
Textile-based cultural heritage is notoriously difficult to digitize. Unlike a stone statue or a flat painting, embroidery is a complex 3D structure composed of twisted silk threads, varying stitch densities, and light-reactive materials like gold and silver thread.
Conventional 2D scanning flattens these nuances, while standard 3D scanning often fails to capture the microscopic relief and "sheen" that define the artifact's character. Furthermore, most digital archives lack the structured metadata (the "how" and "why" of the craft) necessary to train AI models or create truly interactive museum experiences. The authors seek to bridge this gap between physical preservation and digital "creative activation."
Methods
The researchers developed an integrated pipeline moving from physical artifact to procedural digital material:
- Image Acquisition: Using HP Captis optical scanning, they captured high-resolution imagery which was then pre-processed (HDR optimization, color calibration).
- PBR Texture Generation: Using the Adobe Substance 3D suite (Sampler, Designer, Painter), they decomposed scans into core PBR maps: Normal, Height, Roughness, Base Color, and Opacity. This allows the digital asset to react to virtual light exactly like the physical object.
- Procedural Modeling: They created SBSAR (Substance Archive) files. These are not just static images but procedural materials with adjustable parameters like "StitchHeight," "ShineLevel," and "AgeWear," allowing for real-time manipulation in game engines.
- Metadata Structuring: They mapped artifact details to international standards (Dublin Core and Europeana), creating a labeled schema specifically for embroidery (e.g., motif types like "Five Peaks" or techniques like "seed stitch").
Datasets
The study focused on four representative artifacts selected for their diversity in motifs and materials:
- Double Crane Insignia (Gangneung Embroidery Museum)
- Embroidered Dangui (Sookmyung Women’s University Museum)
- Embroidered Pillow End (Gangneung Embroidery Museum)
- Five Peaks Embroidered Silk Panel (Onyang Folk Museum)
The resulting dataset includes 8K reconstructions and a structured labeling schema designed for future AI training in automated classification and semantic search.
Key Insights and Ideas
- Materiality as Data: The "aha moment" here is treating the technique of embroidery as a set of renderable parameters. By isolating the "Height Map," the researchers could digitally replicate the physical relief of a "layered gold stitch," which is essential for the tactile feel of the textile in VR.
- From Archive to Interface: The authors argue that digital heritage shouldn't be a static "bit" in a database. By using SBSAR formats, they enable "Museum 4.0" experiences where a visitor can interactively change a badge's gold thread to silver or swap motifs, fostering a participatory understanding of the craft.
- AI-Ready Heritage: By structuring metadata around specific embroidery techniques (e.g., seed stitch vs. flat stitch), the paper provides a blueprint for training generative AI that "understands" the technical logic of traditional Korean crafts, rather than just mimicking their surface appearance.
Key Results
- High-Fidelity Reconstruction: The team successfully reconstructed complex stitches (like Jaryeonsu) in 3D at 8K resolution, clearly expressing both the fabric grain and the individual thread paths.
- Cross-Platform Interoperability: The assets were proven compatible across Unity, Unreal Engine, and WebXR, demonstrating a scalable path for e-commerce, education, and virtual exhibitions.
- Interactive Parameters: The creation of a "Smart Material" library allows designers to apply "Korean Embroidery" textures to any 3D model while maintaining historical accuracy in how those materials reflect light and show wear.
Limitations and Future Work
The authors acknowledge that their current study is limited to a small sample size of four artifacts. To create a truly robust AI training set, the dataset needs to expand into broader categories like Buddhist and court embroidery.
Future work will focus on:
- Ultra-high-res imaging: Moving beyond current scanning limits to capture even finer thread dynamics.
- AI-driven Curation: Developing AI agents that can automatically describe and categorize textile artifacts based on the established metadata schema.
- Commercial Expansion: Integrating these high-fidelity textures into e-commerce and digital fashion (e.g., CLO 3D) to bring traditional craftsmanship into contemporary consumer markets.
Summary updated