How 99 Lines of Code Let You Create Frozen‑Style Magic with Taichi

Only 99 lines of code let you create ice‑magic effects similar to those in Disney's Frozen, thanks to Taichi, a new computer‑graphics programming language developed by MIT PhD student Hu Yuanming.

Taichi dramatically reduces the barrier to high‑performance physical simulation: a simple scene that takes minutes on a regular PC runs 188× faster than TensorFlow and 13.4× faster than PyTorch, while the code is about one‑tenth the length of traditional implementations.

Installation is as easy as import taichi as ti, similar to TensorFlow. import taichi as ti Hu also provides a complete tutorial, and his work has been accepted at top conferences such as SIGGRAPH 2018, ICRA 2019, and ICLR 2020.

Material Point Method (MPM) and MLS‑MPM

MPM is widely used in visual effects; Frozen used it, but early versions required a week of cluster time for a few seconds of footage. Hu improved it with Moving Least Squares MPM (MLS‑MPM), achieving a two‑fold speedup and enabling new phenomena such as material cutting and bidirectional coupling of rigid bodies.

An 88‑line C++ demo of MLS‑MPM is provided as a reference implementation.

ChainQueen Differentiable Physics Engine

Building on MLS‑MPM, ChainQueen is a real‑time, differentiable hybrid Lagrangian‑Eulerian simulator for deformable objects, delivering high precision in both forward simulation and gradient computation. The research appeared in ICRA 2019 and formed Hu's master thesis.

DiffTaichi

DiffTaichi extends Taichi with differentiable programming, offering ten physical simulators that can be benchmarked against existing baselines. Most models run in 2–3 hours without a GPU; examples include elastic objects, rigid bodies, fluids, light refraction, and elastic collisions.

A differentiable elastic object simulator runs on a 2017 13‑inch MacBook Pro in under ten minutes, while the CUDA version requires 490 lines of code compared to 110 lines in Taichi.

Taichi’s speed rivals CUDA, being 188× faster than TensorFlow and 13.4× faster than PyTorch, and neural‑network controllers typically need only a few dozen gradient steps for optimization.

Why Taichi?

Computer graphics lacks a universal tool like TensorFlow, forcing developers to master low‑level details. Taichi provides a Python‑friendly interface, integrates with IDEs and libraries such as NumPy and Matplotlib, and is distributed as a pip package, lowering the entry barrier for graphics programmers.

Hu Yuanming’s Background

From a top high‑school student to a Tsinghua “Yao Class” graduate, Hu has won multiple programming contests, interned at Microsoft Research Asia, and published several papers in CVPR, SIGGRAPH, and other top venues before completing his MIT PhD.

For more details, see Hu’s Zhihu article, the arXiv paper, and the GitHub repositories.