The Mission
Why Creator DNA?
The end goal isn't just to make educational videos — it's to generate production-quality YouTube videos end-to-end,
adapting to the specific style of any creator using LoRA-style adapters learned from their existing catalog.
But before you can learn a style, you have to measure it.
🎯 The Four DNA Dimensions We're Extracting
Rather than manually designing separate heuristics for scripting, narration, storyboarding, and editing,
the objective is to learn creator-specific style adapters from existing videos so that a strong research document
can be automatically transformed into a compelling creator-style presentation. To achieve this, we extract
each creator's "DNA" across four dimensions:
📖 Narrative DNA
🎙️ Vocal DNA
🎨 Visual DNA
⚙️ Structural / Editing DNA
Total Content Hours by Channel
🏆
Simple History dominates in volume
14,550 shots across 99 videos — nearly 27.7 hours of content. The widest format range in the dataset.
⚡
OverSimplified is the most efficient
Only 32 videos, yet 6,402 shots and 10.9 hours. Averages 200 shots per video — the most shot-dense channel.
🎯
Kurzgesagt: precision output
99 videos, 7,053 shots, the most consistent production cadence. Every video is structurally nearly identical.
Narrative DNA is the information architecture of a video. It's not just what a creator says — it's the
rhythm of when they reveal information, how they open curiosity loops, and how densely they pack ideas per sentence.
The data exposes this through narration rate and speech length distributions.
% of Scenes with Narration (Speech)
Avg Speech Length per Narrated Scene (chars)
Narration Rate — All Channels Ranked
🎤
OverSimplified: Voice-first narrative
87% narration rate — history storytelling demands continuous voice. The script IS the video; visuals annotate dialogue, not the reverse.
🤐
CGP Grey: "Show, don't tell" outlier
Only 61% narration. Grey's narrative DNA includes deliberate silent scenes where visuals argue the point independently — a rare and learnable style signal.
📝
MinuteEarth: Densest sentences
390 chars per narrated scene — nearly 3× the median. Each shot carries a full ecological argument. Long, compound, information-dense sentences are their signature.
⚡
MinutePhysics & CGP Grey: Punchy brevity
Both at ~107 chars/scene. Short declarative sentences, rapid pacing. Their Narrative DNA is "one idea per breath, then cut."
"Narrative DNA is what makes a Kurzgesagt explainer feel different from a CGP Grey essay even when they cover the same topic. One builds tension through visual metaphor; the other spirals inward through argument. The data quantifies this."
— Series analysis note
Vocal DNA sits at the intersection of what's said and how it's said. The dataset lets us proxy this through
speech density per unit of scene time — a creator who fits 390 chars into a 14-second scene is speaking slowly with deliberate pauses,
while one who fits 107 chars into a 3-second scene is at a completely different cadence.
Speech Density: Chars per Second of Scene (Vocal Rate Proxy)
Speech Length Distribution — All Channels (Scatter)
🐇
MinutePhysics: Fastest vocal delivery
At a 2.1s median scene and 107 chars of speech, the implied delivery is rapid-fire. Vocal adapter must learn to compress ideas without losing clarity.
🐢
MinuteEarth: Deliberate, measured delivery
390 chars over 14.7s scenes implies a calm, documentary cadence with natural pauses between clauses — more David Attenborough than Vsauce.
🎭
OverSimplified: Comedic timing signature
87% narration, 88% SFX — the vocal delivery is structured around comedic beats, with audio punctuation (stings, sound effects) that's part of the performance, not decoration.
🎓
TED-Ed: Academic register, moderate pace
115 chars at 8.3s scenes — a clean, neutral academic lecture cadence. Deliberately accessible rather than specialized, matching their cross-curriculum audience.
Visual DNA is the most immediately recognizable layer — you can identify a Kurzgesagt video in one frame. But "style" here isn't
just aesthetics. It's the structural relationship between how long scenes last, how richly they're described, and how tightly
they sync to narration. The data gives us all three dimensions.
Avg Visual Description Length (chars/scene)
Scene Duration Distribution — Full Dataset
Mean Scene Duration by Channel (Editing Tempo = Visual Rhythm)
🌿
MinuteEarth: Richest visual scenes
Average 801 chars of visual description — layered environmental context, organism details, process annotations. Each scene is a paragraph of visual information.
🏃
MinutePhysics: Visual brevity at high speed
Only 539 chars/scene, with 2.1s median cut. Their visual DNA is about reduction — the minimum visual to carry the maximum idea. Hand-drawn whiteboard enforces this.
⚙️
Frame rate is a visual identity signal
Kurzgesagt runs at 60fps exclusively. TED-Ed, MinutePhysics, MinuteEarth all at 24fps (cinematic). Frame rate is part of the visual DNA, not just a technical spec.
🔗
The richness ↔ pacing tradeoff
Slow scenes → rich visuals. Fast scenes → sparse visuals. MinuteEarth is the clearest example: 14.7s scenes, 801 chars. MinutePhysics: 2.1s, 539 chars. Generation must respect this coupling.
Frame Rate by Channel — Production Technical Signature
| Channel |
Primary FPS |
Visual Style Signal |
Structural DNA is the architecture of the video as a whole. Not individual scenes — but the pattern of scenes:
how many shots does this creator use for a 10-minute video? How variable is that across their catalog? What's the
median scene in the middle of their videos vs. the opening? This is the hardest DNA to extract, and the most valuable
for generation — because it controls the macro-level pacing and coherence of the output.
Shots per Video — Min / Median / Max by Channel
Median Video Duration (minutes) — Format Strategy
Speech vs. SFX Presence — Audio Layer Structural Comparison
📐
Kurzgesagt: The most learnable structure
Std dev of 30.2 shots/video — the tightest in the dataset. Their structural DNA is nearly a formula: ~70 shots, ~11 min, 79% narration. The easiest to replicate systematically.
🎲
Simple History: The hardest to pin down
Std dev of 195 shots/video — from 44-shot explainers to 1,592-shot documentaries. Their structural DNA is topic-dependent; generation requires a topic-aware length model.
🎬
OverSimplified: Long-form with high density
Median 18.3 min videos, 200 shots each. Their structural DNA encodes sustained narrative arcs with regular comedic beat breaks — a "chapters + punchline" structure.
⏱️
MinuteEarth & Life Noggin: Micro-format masters
Both median under 4 min. Their structure is tight: single-idea intro → two supporting points → memorable closer. Ideal templates for short-form educational generation.
Cross-DNA Synthesis
Five Style Clusters Emerge
Looking across all four DNA dimensions simultaneously, the 9 channels naturally group into 5 distinct production archetypes.
Each cluster represents a learnable style adapter target — a distinct combination of narrative register, vocal delivery, visual aesthetic, and editing structure.
Radar: Speech vs SFX vs Avg Shot Duration vs Visual Richness — Cluster Mapping
⚡ Rapid Explainer
MinutePhysics · CGP Grey
4–8 min
2–5s cuts
61–75%
Minimal SFX
🧬 Character Science
Amoeba Sisters · Life Noggin
4–10 min
5–12s cuts
68–71%
High SFX
✨ Premium Animation
Kurzgesagt · TED-Ed
5–12 min
6–10s cuts
74–80%
Low SFX
📜 History Storytelling
OverSimplified · Simple History
10–30 min
5–7s cuts
73–87%
High SFX
🌍 Illustrated Nature
MinuteEarth
3–5 min
14s cuts
75%
High SFX
What This Enables
From Analysis to Generation
This DNA analysis establishes the ground truth we'll train against. Each dimension maps directly to a generation component in the pipeline.
📖 Narrative → Script Adapter
Sentence length, information density, and curiosity-gap patterns train the script rewriter that transforms a research document into creator-voice narration.
🎙️ Vocal → TTS LoRA
Speech rate proxies, pause distributions, and prosody patterns train the voice synthesis adapter that delivers narration in the creator's cadence.
🎨 Visual → Image Adapter
Visual description length, scene richness, shot duration, and FPS targets train the image generation pipeline for each creator's aesthetic.
⚙️ Structural → Edit Sequence Model
Shot budgets, video length distributions, and scene-type sequences train the editing model that assembles generated clips into a coherent creator-style video structure.
Part 2: Learning the Adapters
With the DNA extracted, Part 2 will show how these measurements become training targets — building the first
prototype of a creator-style LoRA adapter for script generation. We'll start with the Kurzgesagt cluster:
the most consistent, most learnable, and most data-rich target in the dataset.
Coming Soon → Part 2