Benchmarks

Every performance number on this site comes from a reproducible benchmark you can run on your own hardware.

Latest Results

Measured on Apple M2, 24 GB RAM, macOS. Your numbers will vary by hardware. Run bun run bench:full to measure on your own machine.

OG Engine — Phase Breakdown

Scenario	Text Measure (P50)	Canvas Draw (P50)	PNG Encode (P50)	Full Pipeline (P50)	Full Pipeline (P95)
Baseline (og, 1 line, Outfit)	114µs	50µs	21.39ms	21.57ms	22.79ms
Long text (og, overflow, Outfit)	390µs	78µs	24.34ms	24.83ms	26.41ms
Story format (1080x1920, Outfit)	426µs	98µs	59.37ms	59.96ms	65.02ms
CJK (og, Noto Sans JP)	126µs	79µs	24.12ms	24.34ms	26.92ms

PNG encoding dominates render time. The actual text layout and compositing — the part that replaces Puppeteer’s browser engine — takes under 0.5ms for all scenarios.

Puppeteer Baseline

Scenario	Warm P50	Warm P95	Cold P50	Cold P95
Baseline (og, 1 line, Outfit)	128.75ms	165.98ms	657.55ms	837.24ms
Long text (og, overflow, Outfit)	132.14ms	153.42ms	634.03ms	788.19ms

Warm = browser reused between renders (best case for Puppeteer, typical for long-running servers). Cold = fresh browser launch per render (realistic for serverless / Lambda / edge functions).

Speedup Comparison

Scenario	OG Engine (P50)	Puppeteer Warm (P50)	Speedup
Baseline	21.57ms	128.75ms	6x
Long text	24.83ms	132.14ms	5x

Against cold-start Puppeteer (serverless deployments), the speedup is 30x.

Where the Time Goes

Phase	What it does	Typical time
Text measurement	Word-wrap computation for title + description	~0.1ms
Canvas draw	Gradient, grid, glow, text, decorations composited onto Canvas	~0.05ms
PNG encode	Canvas buffer encoded to PNG binary	~21ms
Full pipeline	All three (= API response time)	~22ms

The speed advantage over Puppeteer comes from architecture, not just raw encoding speed:

No browser launch — zero cold start overhead
No DOM/CSS — text layout computed mathematically, not via Blink
No Xvfb — no headless display server required
10MB vs 200-500MB — per-render memory footprint
500+ concurrent renders — single Node.js process vs 5-10 Chrome instances

Methodology

OG Engine: 1000 iterations per scenario, 50 warmup discarded
Puppeteer warm: 50 iterations, browser reused, 5 warmup
Puppeteer cold: 10 iterations, fresh browser per render
All times measured with performance.now()
P50 (median) used for all headline numbers
Machine info captured in every report

Run It Yourself

git clone https://github.com/phmatray/og-engine
cd og-engine
bun install
bun run fonts:download
bun run bench:full

Results are saved to benchmarks/results/. The markdown report is human-readable; the JSON contains every individual timing for your own analysis.

Source

The benchmark code is in the benchmarks/ directory.