💻 computer science

Adaptive Yarn Cards: Geometry-Aware Curved Texture Proxies for Efficient Rendering of Dynamic Woven Cloth

This paper proposes adaptive yarn cards, a geometry-aware, curved texture-card representation that decouples low-resolution yarn simulation from high-density rendering to efficiently approximate the visual fidelity of dynamic woven cloth while significantly reducing geometric complexity and achieving a 2.4x rendering speedup compared to dense yarn geometry.

Original authors: Jong-Hyun Kim

Published 2026-10-01

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jong-Hyun Kim

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine trying to create a digital version of a piece of woven fabric. To make it look real, you cannot just draw a flat sheet; you must model the thousands of individual threads that make up the cloth, including how they twist, bend, and overlap one another. In the world of computer graphics, this level of detail is called yarn-level modeling. While it produces stunningly realistic images, it is incredibly expensive for a computer to process. Every single thread must be calculated as a separate 3D object, and when the fabric moves, the computer has to re-calculate the position of every thread in every frame. This creates a heavy burden that often slows down animations or makes real-time interaction impossible, especially when the fabric is complex or moving quickly.

The challenge becomes even trickier when the computer needs to simulate the physics of the cloth at a low resolution to save time, but then render it at a high resolution to look good. If you simply fill in the gaps between the simulated threads with more detailed threads for the final picture, the computer ends up doing a massive amount of work to draw geometry that isn't actually involved in the movement. It is like building a full-scale wooden bridge just to paint a picture of it, only to realize you could have painted the picture on a flat board and saved all the wood.

A researcher at Inha University in South Korea has developed a new way to solve this problem, allowing computers to render complex, moving woven fabrics much faster without losing the realistic look of the threads. Their solution, which they call adaptive yarn cards, acts as a smart middle ground between a heavy 3D model and a simple flat image. Instead of building thousands of individual 3D threads for the final picture, the system uses flat, curved pieces of texture—like playing cards made of fabric—that bend and twist to follow the shape of the cloth. These cards are not just flat squares; they are shaped to match the curves of the fabric as it moves, and they carry a detailed picture of the woven pattern on their surface.

The brilliance of this method lies in how it decides where to use these cards and how big they should be. The researcher realized that a single size of card does not work for every part of a moving cloth. In areas where the fabric is hanging straight or moving gently, a large card is perfectly fine and efficient. However, in areas where the fabric is crumpled, twisted, or hitting an object, a large card would look blocky and inaccurate. To handle this, the system constantly measures the shape of the fabric. If it detects a sharp bend or a complex twist, it automatically breaks that area down into smaller, more numerous cards to capture the detail. If the fabric is smooth, it keeps the cards large to save computing power. This process happens automatically and continuously as the cloth moves, ensuring that the computer only spends its energy on the parts of the fabric that need it most.

To test their idea, the researcher ran a series of simulations involving various scenarios, such as a curtain draping under gravity, a piece of cloth being twisted, fabric shrinking due to heat, and cloth colliding with a sphere. In every case, they compared their new method against the traditional approach of rendering every single thread explicitly. The results showed that the adaptive card system could produce images that looked nearly identical to the heavy, detailed models but required significantly less work. On average, the new method was about 2.4 times faster than the traditional approach. More importantly, it was much more efficient than using cards of a fixed size, which either wasted resources on simple areas or failed to capture detail in complex ones.

The researcher found that this approach works well even for fabrics that are not flat, such as cylindrical sleeves or conical skirts, which start with complex 3D shapes and then crumple or fold. The system successfully adapted the size and number of cards to match the changing complexity of these shapes. While the method is not perfect for extreme close-ups where a viewer might need to see the individual cross-section of a thread, it excels at the distances where most people view digital cloth. By decoupling the physics simulation from the visual rendering, the researcher has created a way to make digital fabrics look dense and realistic without forcing the computer to do the heavy lifting of modeling every single thread. This advancement suggests a future where complex, realistic cloth animations can run smoothly on standard computers, opening the door for more immersive virtual environments and interactive simulations.

Technical Summary: Adaptive Yarn Cards

Problem Statement

Yarn-level cloth representations offer high visual fidelity by explicitly modeling woven structures and the geometric characteristics of individual yarns. However, they incur substantial computational and geometric costs because rendering requires the explicit generation and processing of a large amount of high-resolution yarn geometry. This issue is exacerbated when the resolution required for visual rendering exceeds that of the physical simulation. In such cases, interpolating dense rendering yarns between sparse simulation nodes generates complex geometry that does not contribute to the underlying dynamics, significantly increasing the rendering load. While hair rendering techniques often approximate strands using textured cards, these methods typically rely on planar quads or fixed-size cards, which fail to effectively approximate the complex, doubly-curved surfaces and intersecting warp-weft structures of dynamically deforming woven cloth.

Methodology

The paper proposes Adaptive Yarn Cards, a novel representation that decouples low-resolution yarn simulation from high-density rendering. The method replaces explicit high-density yarn geometry with curved, geometry-aware textured proxies.

Core Pipeline

Decoupling Simulation and Rendering: The system uses a low-resolution yarn structure ( $Y^S$ ) for physical simulation. High-density visual appearance is not generated as geometry but is compressed into a procedural texture ( $T_{yarn}$ ) defined over the cloth's 2D parameter domain. This texture encodes the periodic warp and weft patterns, including over-under relationships at intersections.
Curved Textured Card Representation: Instead of planar billboards, each card is a curved mesh constructed by sampling the current simulated cloth surface ( $S_t$ ). Vertex positions are defined as $p_k(u,v,t) = S_t(u,v) + \delta n_t(u,v)$ , where $\delta$ is a small offset to prevent z-fighting. Cards preserve global UV coordinates to ensure the yarn pattern remains continuous across card boundaries, regardless of subdivision.
Geometry-Aware Adaptive Partitioning: To avoid the inefficiency of fixed-size cards (which over-approximate flat regions and under-approximate complex ones), the method employs a recursive subdivision strategy based on a geometry-aware error metric.
- Error Metric: The error ( $E_k$ $E_{k}$ ) for a candidate region combines three normalized terms:
  - Normal Variation ( $\tilde{E}^N$ ): Measures broad-scale surface orientation changes.
  - Local Curvature Variation ( $\tilde{E}^C$ ): Measures local bending via differences between adjacent surface normals.
  - Bilinear Geometric Deviation ( $\tilde{E}^G$ ): Measures the deviation of the surface from a simple four-corner bilinear patch to detect twisting and nonlinear deformation.
- Adaptive Logic: If the combined error exceeds a split threshold ( $\tau_{split}$ ), the region is subdivided. If the error falls below a merge threshold ( $\tau_{merge}$ ), child regions are merged. A hysteresis interval between these thresholds prevents temporal popping during dynamic deformation.
Temporal Coherence: The adaptive hierarchy is updated every $N_{update}$ frames (default 6) rather than every frame to balance responsiveness with computational cost.

Key Contributions

Adaptive Yarn Cards Representation: A method that decouples simulation resolution from rendering density, replacing explicit high-density yarn geometry with curved textured proxies that follow dynamic cloth deformation.
Curved Textured-Card Rendering: A technique that compresses woven appearance into texture space while using cards that deform as curved proxies sampled from the simulated surface, rather than as static planar quads.
Geometry-Aware Error Metric: A composite metric combining normal variation, local curvature, and bilinear geometric deviation to drive an adaptive partitioning scheme that allocates geometric resources based on local complexity.
Comprehensive Evaluation: A comparison of the proposed method against dense yarn geometry, fixed-size curved cards, and textured simulation meshes across various scenes (draping, twisting, thermal shrinkage, collision, and non-planar configurations).

Experimental Results

Experiments were conducted on scenes including draping, twisting, thermal shrinkage, sphere collision, and complex non-planar structures (conical skirts, cylindrical sleeves).

Geometric Complexity: The adaptive method substantially reduces the number of rendering primitives compared to dense yarn geometry. On average, it reduces the card count by approximately 59.5% compared to a fixed-size card configuration (e.g., 91.1 adaptive cards vs. 225 fixed cards).
Rendering Performance: The proposed method achieves an average rendering speedup of approximately 2.4× over explicit dense yarn geometry. It outperforms fixed-size cards by allocating resources only where geometric complexity demands them.
Visual Quality: The method preserves the major woven appearance and overall shape of the dense-yarn baseline. While it does not recover individual yarn cross-sections or deep self-occlusion at close range, it maintains visual continuity and pattern fidelity at mid-to-far viewing distances.
Ablation Studies: Removing any of the three error terms (Normal, Curvature, or Geometric Deviation) resulted in higher visual error (LPIPS) and silhouette errors, confirming that the terms complementarily capture different spatial scales of deformation.
Sensitivity: The split threshold ( $\tau_{split}$ ) and update interval ( $N_{update}$ ) allow for tuning the trade-off between visual fidelity, temporal stability, and rendering performance.

Significance and Limitations

The paper claims that Adaptive Yarn Cards provide a lightweight representation suitable for real-time and interactive rendering of complex deformable cloth. By separating the visual density of the weave pattern from the geometric complexity of the surface, the method allows for high-fidelity rendering without increasing simulation resolution.

The author explicitly states that this method is not a complete substitute for explicit yarn-level rendering at all scales. It is most effective for mid- to far-range rendering where macroscopic deformation and weave organization dominate over individual yarn cross-sections and parallax. The method assumes a consistent cloth parameter domain and does not natively handle topology changes like tearing or seams. Future work is suggested to integrate view- and screen-space criteria and hybrid Level-of-Detail (LOD) transitions for close-range rendering.