Manimgl video

ManimGL Video

ManimGL is 3Blue1Brown's personal rendering engine. It is not Manim Community Edition. Both are called "Manim", both define a class named Tex, and the two mean opposite things. Confirm which engine the project uses before writing a line.

Detect the engine from the project:

• from manimlib import * / manimgl CLI / custom_config.yml → ManimGL, this skill
• from manim import * / manim -ql CLI / manim.cfg → Manim CE, stop — this skill does not apply

Do not confuse with Manim CE

Every row below was verified against manimlib source. The Tex row is the dangerous one: the name exists in both engines, so the wrong choice raises no error — it silently renders text as math or math as text.

	ManimGL (this skill)	Manim CE (different tool)
install	pip install manimgl	pip install manim
import	from manimlib import *	from manim import *
CLI	manimgl file.py Scene -w -l	manim -ql file.py Scene
math mode	Tex("x^2")	MathTex("x^2")
text mode (LaTeX)	TexText("hello")	Tex("hello")
draw a shape	ShowCreation(mob)	Create(mob)
default background	#333333	#000000

These CE names do not exist in manimlib — using them is an immediate NameError: Create, MathTex, Unwrite, AddTextLetterByLetter, Angle, RightAngle, Star, Axes3D.

Minimal scene

from manimlib import *


class Demo(Scene):
    def construct(self):
        title = Text("Pythagoras", font_size=48)
        eq = Tex(R"a^2 + b^2 = c^2", font_size=60)
        eq.next_to(title, DOWN, buff=LARGE_BUFF)

        self.play(Write(title))
        self.play(FadeIn(eq, shift=UP))
        self.wait()

manimgl file.py Demo        # interactive window, no file written
manimgl file.py Demo -w -l  # write low-quality mp4

Every scene subclasses Scene, InteractiveScene, or ThreeDScene and defines construct(self). 3b1b's own video code overwhelmingly uses InteractiveScene (it adds mouse selection and editing in the preview window); plain Scene is correct for headless rendering and is the safer default when writing code for someone else to render.

Workflow

1. Write the scene outline in prose first — one thing proved per scene. Do not open an editor until the beats are decided.
2. Run manimgl file.py Scene with no -w. The interactive window is the fast feedback loop; writing a file on every iteration wastes minutes.
3. Add self.embed() at the point of interest to drop into an IPython shell with the scene live. Inspect and nudge mobjects there instead of re-rendering.
4. Once composition is settled, -w -l for a low-quality file to judge timing.
5. Fixing a late beat? -n <index> starts at animation N instead of replaying everything from the top.
6. Only after timing is final, render --hd or --uhd.

Tighten typography, color, and spacing after motion works — not before.

On-screen text

Default to English. Labels, titles and annotations are English unless the user asks for another language, or the surrounding project is already in one. Conversing in another language is not by itself such a request — ask if unsure. Non-Latin scripts additionally need the font and animation care described in references/api-core.md.

The animation carries the argument. A sentence that narrates what the viewer is already watching competes with the visual for attention and duplicates whatever a narrator would say over it.

The test: if a narrator could speak the line, it does not belong on screen.

Cut — prose that restates the motion:

• "Unroll it into one row in reading order" over a grid that is visibly unrolling
• "Without this, shuffling the patches changes nothing"
• "After 12 blocks it has gathered information from every patch"

Keep — anything the picture cannot state on its own:

• labels and names: axis titles, LayerNorm, Q, K, V
• counts, units, and specs: 4×4 = 16 patches, 224×224 → 196 patches of 16×16
• formulas: Attn = softmax(QK^T/√d)V
• emphasis: color coding, a highlighted element, Indicate, FlashAround

Emphasis is not narration — keep every bit of it. Marking one patch red and tracking it across scenes explains more than a caption would.

Prefer noun phrases to sentences, including for section titles: 1. Patch split over 1. The image is cut into patches. Where a scene seems to need a sentence, the usual fix is a clearer visual, not a longer caption.

Reference files

Open only what the current task needs; do not preload all four.

File	Open when
references/api-core.md	Tex/Text, geometry shapes, VGroup, positioning, colors
references/api-graphs.md	Axes, NumberPlane, function graphs, ValueTracker, updaters
references/animations.md	Which animation class to use, timing, lag_ratio, rate functions
references/cli-config.md	CLI flags, custom_config.yml, reproducing the 3b1b look

Runnable starting points, verified against manimlib source:

• assets/smoke_test.py — geometry and Text only, no LaTeX required. Use this first to prove the install works.
• assets/equation_graph.py — Tex + Axes + ValueTracker. Requires a working LaTeX install.

Footguns

Tex takes multiple strings. Tex("a", "+", "b") splits into indexable submobjects. To reliably address a substring, pass isolate=[...] and retrieve with get_part_by_tex / get_parts_by_tex, or color via t2c={...}. Indexing a Tex by raw integer position breaks the moment the LaTeX changes.

LaTeX is required for Tex, TexText, and Brace. Text and Code render through Pango and need no LaTeX. When LaTeX is unavailable, a scene can still be built entirely from Text plus geometry — say so rather than letting the render fail.

ShowCreation, not Create. The most common CE reflex, and it fails loudly.

Never Write CJK text. Write traces glyph outlines, and a Hangul or Han syllable is several independent contours, so every in-progress glyph renders as disconnected strokes — illegible for the whole animation. Changing the font does not help: regular and thin weights mangle identically. Use FadeIn, or AddTextWordByWord when progressive reveal is wanted. Keep Write for Latin text and Tex.

Updaters must be removed. A mobject with add_updater keeps mutating for the rest of the scene. Call clear_updaters() when the effect is over, or the object will drift during later beats.

always_redraw(func) rebuilds the mobject every frame. It is the correct tool for something whose shape depends on a tracker, and the wrong tool for something that merely moves — use add_updater or f_always for that. Rebuilding complex mobjects each frame is the usual cause of a slow preview.

self.play() needs an animation, not a mobject. To animate a method call use the .animate syntax: self.play(square.animate.shift(RIGHT)).

Background is #333333 by default, noticeably lighter than CE's black. 3b1b's production config overrides it to #000000; see references/cli-config.md.

Attribution and licensing

• 3b1b/manim (the engine) is MIT. Deriving code from manimlib and example_scenes.py is unrestricted.
• 3b1b/videos (the per-video scene code) is CC BY-NC-SA 4.0 — non-commercial and share-alike. Do not copy scenes from it into user projects without telling the user about the license. It also depends on manim_imports_ext.py and the repo's custom/ package (pi creatures, backdrops), so it does not run standalone.

Write fresh code modeled on the API. Do not paste from 3b1b/videos.