Database Sketches · Postgres

Postgres bloat, drawn in 33 seconds

An influencer-style explainer video, made with AI.

  • 33 sfinal cut
  • 96 sof drawing, at 3×
  • 4chained clips
  • 1unbroken voice take
0

The reference

The goal was the feel of sketch videos like this one: a camera locked overhead, a real hand drawing one dense page, sped up, with tool changes cut out and a voice that never stops.

Reference · 15 s slice“Man City Charges explained”, by @_butilebo (credit: @_sportsball)
1

Started with a floating pen

the foundation

The sketch page writes every label stroke by stroke with single-line fonts and shades with pencil hatching, on a clock we control frame by frame. It was exact, but a floating pen isn't a person.

Floating penthe live page, rendered
2

A 3D hand in the browser

looked like 2000s CGI

We built a rigged Blender hand, then photo-derived 3D scans of a generated hand, rendered live with three.js. The motion was ours, but it never stopped looking rendered, and the scanned pencil fused into the fingers.

Scanned 3D handthree.js over the SVG page
3

Let AI draw everything

pretty, but it cheats

Given a first and last frame, Seedance invented the drawing in between. The hand looked real, but lines faded in instead of being drawn, coloring ran at superhuman speed, and quality fell apart over longer takes.

First + last frameSeedance 2.5 image-to-video
Lettering testlegible, but far too slow to write out
4

Engine draws, AI adds the hand

the breakthrough

We rendered our engine's drawing as video and asked video-to-video models to replace the floating pen with a real hand. We tried Kling O3, MiniMax H3 and Seedance 2.5. Seedance traced our video frame for frame, so the ink stayed exactly where and when our engine put it.

Inputour engine
Seedance 2.5 editthe pick
Kling O3 editclose, but tints a bar
5

Teach the engine to move like a hand

human timing

Because the model copies our timing, our timing had to be human. We let Seedance draw the bars freely at its own pace, measured that take (about 1.2 s per outline, slowing at corners, 1 s pauses, 3 s tool swaps), and copied those numbers into a film mode: hand-speed shading, no instant strokes, visible pen lifts, and tools set down and picked up at the desk.

AI drawing freelywhat we measured
Film modeour engine, human timing
After Seedancereal hand, real swaps
6

Speed it up, cut the swaps

the format

Like the reference, we film at human speed and play it back fast. Real motion sped up still reads as human. Tool changes are jump-cut: the hand leaves with the pen and comes back drawing in blue. Four clips, about 96 s of drawing, play at 3×.

First timelapse2×, swaps cut
7

Add narration

ElevenLabs v3 · Eric

Stitching separate lines sounded jumpy. The final narration is one continuous take, left unedited, and the video bends to fit it instead: each section plays between 2.6× and 3.5× so “blue is live data” lands on the blue, “autovacuum” on the vacuum notes, and “that’s bloat” on the arrow. Captions light up word by word from the voice's own timestamps.

Beforestitched lines
Afterone take, video retimed
✓

The recipe

  1. Draw it in the engine, in film mode: human-measured timing, outlines first, colors after, tools swapped at the desk.
  2. Render ≤30 s clips with a fixed overhead camera and a short pre-roll where the tool comes in.
  3. Seedance 2.5 edit each clip: “every floating tool is held by a real hand…”, with a reference photo of the hand.
  4. Cut the swaps, join, speed up about 3×.
  5. Write the script for the fast cut, generate it as one take, and retime the video sections to the voice.
  6. Captions from the voice's word timestamps, burned in.