Swish: Neural Network Cloth Simulation on Madden NFL 21

How to read this

Category

Method / production: real-time neural cloth for games

Contributions

• A neural network predicting cloth shape from a character's skeleton pose at real-time rates, shipped in Madden NFL 21
• Represents detail with both mesh deformations and a database of normal maps to capture high-resolution wrinkles
• Shares implementation and performance details for introducing ML into real-time character pipelines, targeting tight clothing where traditional real-time cloth struggles

Context

Builds on learning-based clothing animation for virtual try-on (Santesteban et al.) and adapts it to a shipped real-time game setting, billed as the first neural cloth simulation in a shipped game.Builds on: Learning-Based Animation of Clothing for Virtual Try-On

Correctness

Demonstrated on NFL player jerseys (tight clothing driven by skeleton pose), so results are validated in that domain; the pose-to-shape mapping and normal-map database are tailored to that case, and generalization to loose or free-flowing garments is not the target.

Clarity

Practitioner-oriented and accessible; a first pass conveys the architecture and the mesh-plus-normal-map representation.

How to read it

First pass for the representation and shipping constraints; revisit the implementation and performance section if you are building a real-time character ML pipeline of your own.