Watch any digital garment in a game, film or virtual try-on and you’ll see the same thing. Cloth folds nicely while it moves, then springs back to perfectly smooth when it stops. Real clothes don’t. Linen stays crumpled, jeans hold creases behind the knee, and a shirt remembers the elbow you bent all day. Those marks carry a lot of information: what the fabric is, how the garment has been worn, even who wears it.
Learnable Persistent Wrinkle Formation in Cloth Simulation, by Deshan Gong, Ningtao Mao, Xiaoyuan Yang, Xinyu Lu, He Wang and Taku Komura, accepted to SIGGRAPH Asia 2026 (arXiv, Sept 12), addresses that gap directly.
The problem is data, not just math
Cloth simulators can already model plasticity, meaning deformation that doesn’t fully recover. What nobody had was a reliable way to set those parameters for real fabrics. No precise dataset existed, and no method to estimate the values from one. So artists tuned wrinkle behavior by eye, fabric by fabric, if they bothered at all.
Fabric-101
The team’s first contribution is Fabric-101, a dataset of more than 101 common fabrics measured under textile-industry standards. The key is how they measured. Each fabric was stretched and released repeatedly (cyclic loading and unloading), which separates three different behaviors:
- Elastic: deformation that recovers on its own
- Recoverable: friction-driven deformation between fibers that can be undone, the kind you shake or steam out
- Plastic: deformation that doesn’t come back
Existing cloth datasets blend these together. Pulling them apart is what makes persistent wrinkles possible to model.
A simulator that learns from the measurements
The second contribution is a differentiable cloth simulator that combines an elasto-plastic model with friction. Because it’s differentiable, it can work backward (using the adjoint method) from each fabric’s measured stretch-and-release curve to the physical parameters that reproduce it. Each fabric gets its own learned wrinkle behavior, taken from real material data instead of hand tuning.
The authors report that simulated garments form persistent wrinkles that look and behave like the real fabrics across many materials and motions.
Why creative practitioners should care
- Digital fashion and virtual try-on. Showing how a garment wears, not just how it drapes on a mannequin, is the difference between a product image and a believable one.
- Character animation and film. Costume continuity (a jacket that’s visibly been lived in) is currently painted in or sculpted by hand.
- Games and avatars. Clothing that picks up its wearer’s habits over a session is a new kind of expressive detail.
It will take time to reach production tools. But because the code and dataset are public, simulator developers and researchers can start testing it now, and that’s usually how a SIGGRAPH idea ends up in software people actually use. Co-author Taku Komura’s long record in character animation research makes that path more likely than usual.