The third path for AI educational video.

Not a face on screen. Not pixels guessed by a model. Concepts rendered from source — formulas, diagrams, code — drawn deterministically. Built for teaching, not broadcasting.

Try X-Pilot Free How rendering works

scene · faraday.tsx1920×1080 · 60fps

source

maxwell_em.pdf · §3.2

VERIFIED

Faraday’s law

∇ × E = −∂B/∂t

LATEX→MATHJAX→FRAME

FRAME 0428 / 1800 SHA a3f9·c70d·…·e824

diagram · mermaid

scene.tsx

export const Scene = () => (
  <MathJax tex={"\nabla\times E"} />
);

The landscape

Three paths. One built for teaching.

Avatar reads a script. Generative samples pixels. Knowledge visualization renders from source.

Path 1

Avatar-led video

Synthesia, HeyGen, Colossyan. A synthetic face reads your script.

BestMarketing, town halls, CEO messages.

WeakTeaching — presenter steals attention from content.

RUN ×3 · DRIFT

SAMPLED

~12% drift

Path 2

Generative AI video

Sora, Veo, ChatGPT video. The model guesses pixels.

BestReels, teasers, cinematic shorts.

Weak10–15% visual hallucination. Drifts every render.

E = mc2

DETERMINISTIC

SHA · a3f9

physics.pdf · §1.40% drift · byte-identical

Path 3 — X-Pilot

Knowledge visualization

Concepts rendered from source. No avatar. No guessing.

BestSTEM, exam prep, medical, compliance.

Edge0% hallucination · audit-traceable · reproducible.

What gets rendered.

Six primitives, each backed by a typed renderer. No model invents the pixels.

latex · mathjax

x=

−b±√b2−4ac

2a

EXACT

Formulas & equations

LaTeX in. MathJax out. Every subscript, every Greek letter, exact.

mermaid · auto-layout

ABCD

Diagrams & flowcharts

Explicit nodes, explicit edges. Molecules, circuits, decision trees.

Code & technical snippets

1  def binary_search(arr, x):
2      lo, hi = 0, len(arr)
3      while lo < hi:
4          mid = (lo + hi) // 2
5          if arr[mid] == x: return mid

type-aware highlightline-by-line reveal

Data & charts

Bars, lines, scatter, timeline. Animated to build intuition.

Annotated text

Regulations and standards quoted verbatim, walked clause by clause.

Animated mechanisms

Reactions, process flows, system architectures — on a frame clock.

Who it’s for

Four jobs, one shape.

Domains where the content matters more than the presenter.

JTBD-1\ STEM teaching \ Chemistry, physics, biology, math. JTBD-2\ Exam prep \ IGCSE, JEE, NEET, USMLE, COLREG. Letter-accurate. JTBD-3\ Medical & technical \ Clinical procedures, engineering training. JTBD-4\ Compliance & SOP \ HIPAA, regulations. Audit-traceable.

Scored on what matters for teaching.

The pipeline

Document → frame. Five steps. Zero sampling.

Same input, same output. Always.

source.pdf

E=mc²

01 · Parse

{
  "type": "formula",
  "latex":
    "E=mc^2",
  "src": "p.12"
}

02 · Map

scene.tsx

const S=()=>
 <Math
  tex={x}/>

03 · Script

sandbox

• remotion4.0.278
• react18.3.1
• mathjax3.2.2

LOCKED

04 · Render

E=mc²

EXACT

05 · Verify

1. Parse

Source → structured outline

Formulas keep their LaTeX. Diagrams keep their nodes. Code keeps its tokens. Nothing paraphrased.

Concept → typed renderer

MathJax for formulas. Mermaid for flowcharts. Shiki for code. D3 for data. Rule-based, reproducible.

Editable narration, line by line

Edit through natural-language editing. Re-render the affected segment only.

Remotion in a sandboxed pipeline

Pixels come from code, not a sampling model. Details on the deterministic rendering page.

Byte-identical — every formula traced to a passage

Re-run the same input and the output is byte-identical. Audit-grade. See why accuracy matters.

The pedagogy

The screen should be the concept.

Three decades of multimedia learning research, summarised.

Mayer, 2001+

Multimedia & coherence principles

Words plus relevant pictures = better learning. Extraneous on-screen elements degrade it. A talking head is extraneous when the subject is a reaction or a code block.

Guo et al., 2014 · edX

6.9M MOOC sessions analysed

Khan-style tablet drawings sustained higher engagement than studio talking-head footage. Replicated across chemistry, math, CS.

3Blue1Brown · Khan Academy

The camera never lingers on the presenter

The two best-rated technical channels share one trait: animations and worked examples occupy the frame. Knowledge visualization is that pedagogy, productized.

High-stakes subjects

One wrong character invalidates the lesson

Exam prep, clinical training, regulatory compliance. Probabilistic visuals violate that requirement. Deterministic rendering does not.

Honest scoping

When this isn’t the right fit.

Knowledge visualization is for teaching. Pick a different tool when…

• The human face is the message. CEO town halls, brand spots. Use an avatar tool.
• You need vibe, not accuracy. Cinematic shorts, teasers. Use generative video.
• It’s a software walkthrough. Live screen recording wins for support reels.
• There is no source document. Capture the knowledge first. Then visualize.

Go deeper.

Why Accurate \ Why accuracy is the educator’s lifeline. No Talking-Head Avatars \ Why virtual presenters harm learning outcomes. Deterministic Rendering \ How Remotion and sandboxes replace generative guessing. Long-Form Guide \ A 4,000-word deep dive into knowledge visualization for educators. X-Pilot vs Synthesia \ Knowledge visualization vs avatar-led video, side by side. Category Analysis \ How knowledge visualization sits between avatar and generative AI.