How 99 Lines of Code Let You Create Frozen Magic with Taichi’s Differentiable Physics

99 Lines of Code for Ice Magic

With only 99 lines of Taichi code you can generate ice‑snow effects similar to those in Disney's Frozen , achieving performance that is 188× faster than TensorFlow and 13.4× faster than PyTorch on a regular PC. import taichi as ti The language is easy to install like TensorFlow and comes with a complete tutorial written by its creator.

88 Lines of Code Simulating Real Physics

The core technique is the Material Point Method (MPM), widely used in visual effects such as Frozen . Hu improved MPM by integrating Moving Least Squares (MLS) to create MLS‑MPM, which doubles the simulation speed and enables new phenomena like material cutting and rigid‑body coupling.

His MLS‑MPM demo, implemented in just 88 lines of C++, was published at SIGGRAPH 2018 and became a reference implementation for learning MPM.

ChainQueen Differentiable Physics Engine

Building on MLS‑MPM, Hu derived gradients of the simulation, allowing gradient‑descent optimization of neural‑network controllers. This led to the ChainQueen engine, a real‑time, differentiable hybrid Lagrangian‑Eulerian simulator for deformable objects, presented at ICRA 2019.

DiffTaichi

DiffTaichi extends Taichi with differentiable programming, enabling the creation of ten physics simulators that can be benchmarked against existing baselines. Most models run in 2–3 hours without a GPU, and examples include elastic objects, rigid bodies, fluids, light refraction, and even a billiards scene.

Compared to hand‑written CUDA, the differentiable elastic object simulator uses only 110 lines of code versus 490 lines, while maintaining comparable speed.

Why Taichi?

Computer graphics historically lacked a TensorFlow‑like universal tool, making development difficult. Taichi fills this gap by offering a Python‑friendly API, seamless integration with NumPy and Matplotlib, and easy installation via PyPI, thereby lowering the barrier for graphics research.

Hu Yuanming’s contributions have been accepted at SIGGRAPH 2018, ICRA 2019, NeurIPS 2019, and ICLR 2020, earning high praise from leading graphics scholars.