ManimGX renders Manim scenes 90x faster using Rust and wgpu

A thirty six second math or product animation written in Manim Community Edition takes 278.83 seconds to render on an Apple laptop, long enough that an agent writing one scene at a time is waiting on the render loop instead of the writer. ManimGX, a new Rust rewrite of the same Python API published seven days ago by Academa, renders the same scenes in 3.08 seconds on the same hardware.

The project is at academa-labs/manimgx under the MIT license, with 732 stars and 47 forks on day seven. The native Python wheels are published to PyPI and a browser build is on npm. The authors are Abdullah Geduk and Sina Atalay at Academa, a company whose homepage describes animated math and 3D explainers for school and research.

The benchmark

The repository publishes the measurement in full: five scenes at 1080p60, total output 36.2 seconds of video, rendered on an Apple M4 Pro with the fastest encoding preset each engine offers. ManimGX finished in 3.08 seconds. ManimGL took 35.28 seconds, Blender Workbench 57.61, ManimCE 278.83, and Blender EEVEE 601.40. The 90.6 times gap over ManimCE and 11.5 times over ManimGL are the project's own figures, measured against the project's own scenes, and the scripts, settings and runs are in the benchmarks folder for anyone who wants to repeat them. The authors show the same frame rendered by each engine beside the chart, so a reader can see what the engines produce at matched quality rather than only how long each one took.

What is under the Rust

The renderer uses wgpu for native GPU work and WebGPU in the browser. The 3D lighting is adapted from Google's Filament: physically based materials, image-based lighting, ambient occlusion, bloom, and AgX tone mapping. Text and math are typeset with Typst, and LaTeX input is translated to Typst by mitex, so there is no system LaTeX install to manage. The native build encodes H.264 with x264 and decodes imported audio through FFmpeg; a browser build runs Python scenes inside a Pyodide worker and renders through WebGPU with no server. On Linux with no suitable GPU, the wheel bundles Mesa's lavapipe for Vulkan on the CPU.

Where agents fit

The README calls ManimGX "an API for 3D video in the age of agents". The library is fully typed, including keyword arguments and the .animate chain, so the editor hints a human reads are also the hints a coding agent reads. A manimgx inspect command produces storyboard frames and 2D layout checks at play endings or at times chosen in the scene file, which gives a scene-writing agent a cheap feedback loop before a full render. The project also publishes an llms.txt setup file so an agent in a coding IDE can install the library and render a first animation from a one-sentence prompt.

A team already using Manim CE for math or product explainers can run the same scenes through ManimGX without rewriting them, and a team that was priced out of motion design because each iteration cost a long render now has one that returns in seconds. The native wheels are GPL-3.0-or-later as a whole, because they bundle x264, so a company shipping ManimGX inside a closed product has to use the browser build or compile without x264.

Source

academa-labs/manimgx on GitHub, the documentation and gallery at manimgx.academa.ai, and the benchmark folder with the five scenes and the matched-quality comparison frames.