Abstract
Traditional cloth simulation relies on discrete meshes, which often struggle to balance computational efficiency with the fine-grained detail required for realistic wrinkles. This paper presents a novel approach using Implicit Neural Representations (INRs) combined with a multi-resolution grid encoding and a minimax adversarial sampling strategy. By representing cloth as a continuous function rather than a fixed grid, the method allows for mesh-free simulation that adaptively focuses computational resources on high-frequency details. Experimental results demonstrate a 346x speedup in training time and superior visual quality compared to traditional mesh-based and standard MLP approaches, particularly in complex interaction scenarios.
Introduction
In computer graphics, representing realistic cloth is a long-standing challenge. Traditional methods rely on discrete meshes—grids of vertices and edges. To capture fine, high-frequency wrinkles, these meshes must be extremely dense, which leads to prohibitive computational and memory costs. Conversely, low-resolution meshes suffer from "discretization artifacts," where wrinkles appear jagged or fail to form entirely.
While machine learning has attempted to address these issues, current supervised models require massive amounts of expensive ground-truth data. Unsupervised approaches often remain tethered to the underlying mesh structure, inheriting its resolution limitations and suffering from slow convergence. This research proposes a shift toward continuous representations, asking if cloth can be simulated as a function that "learns" where to place detail.
Methodology
The authors introduce three core innovations to move beyond the limitations of discrete meshes:
1. Multi-resolution Grid Encoding
Instead of a standard Multi-Layer Perceptron (MLP), the system utilizes a hierarchical grid encoding. This structure allows the network to capture broad, low-resolution shapes and tiny, high-resolution wrinkles simultaneously and efficiently. This architecture is inspired by neural radiance fields (NeRFs) but is specifically optimized for 3D surface deformation.
2. Mesh-Free Local Sampling
Rather than calculating physics on fixed mesh vertices, the system samples points in the cloth's 2D "UV space" (its flat coordinate system). For each sampled point, it constructs a temporary "local sampling structure"—a small patch of triangles—to calculate physical properties like strain (stretching) and bend (folding). This allows the simulation to be truly continuous.
3. Minimax Adversarial Training
To achieve spatial adaptivity, the simulation is framed as a minimax game between two components:
- The Sampler: Acts as an adversary that "hunts" for regions where the physical error (loss) is highest—typically where complex wrinkles should form.
- The Simulator: Adjusts the neural network weights to minimize the error identified by the sampler. This ensures the model spends its computational budget on complex folds rather than flat, uniform areas of the fabric.
Experimental Evaluation
The model was evaluated on several challenging interaction scenarios, including standard primitives (balls, tori) and complex characters like a "Scorpion" and an "Ogre." Because the approach is unsupervised and physics-based, it does not require external datasets; instead, it minimizes the cloth's potential energy (strain, bend, and gravity) as its primary loss function.
Key Results
- Efficiency: The multi-resolution model achieved a 346x speedup in training time compared to a baseline MLP. Specifically, it converged in just 28 seconds (400 epochs), while the standard MLP required over 2.5 hours (500,000 epochs) to reach a comparable state.
- Visual Quality: In head-to-head tests against traditional meshes with an equivalent number of free variables, the INR method produced symmetric, detailed wrinkles. In contrast, the mesh-based versions produced jagged artifacts or failed to capture the wrinkles entirely.
- Convergence: The adversarial sampling strategy significantly reduced the number of epochs needed for convergence by prioritizing "difficult" regions of the cloth.
Discussion and Limitations
The power of this approach lies in its continuity; wrinkles are not "locked" to mesh edges, resulting in more natural patterns. Furthermore, the random orientation of local sampling structures allows the model to handle anisotropic materials—fabrics that stretch differently in different directions—more effectively than fixed-grid methods.
However, some limitations remain. The current model is restricted to a square UV domain, making it difficult to represent complex garments with irregular boundaries, seams, or holes. Additionally, physical loss weights (e.g., bending vs. stretching) currently require manual tuning for different material types. Future work aims to extend this framework to real-time applications such as virtual try-ons and high-fidelity character animation for film.
Summary updated