Build a benchmark chart for your article
Configure an accuracy-vs-parameters chart, pick a theme, and copy a snippet you can drop into any blog post, docs page or README. The chart renders live from Vision Analysis data and updates automatically as new benchmarks land.
Picked models are drawn in full colour with labels. Everything else stays as grey context dots. Pick a whole family to draw its scaling curve.
Paper-reported mAP covers every model. Pick a hardware + runtime to plot measured accuracy instead (only models benchmarked on that setup will appear).
System follows each visitor's OS light/dark setting automatically.
The chart fills its container and keeps an 8:5 ratio, so its height follows the width automatically with no scrollbar. Cap the width if you do not want it to stretch across a wide column.
<div style="position:relative;width:100%;padding-top:62.5%">
<iframe
src="https://visionanalysis.org/embed/scatter?highlight=yolonas-s%2Cyolonas-m%2Cyolonas-l"
style="position:absolute;top:0;left:0;width:100%;height:100%;border:0;border-radius:12px"
loading="lazy"
title="Accuracy vs parameters - Vision Analysis">
</iframe>
</div>How it works
The widget is a plain URL under /embed/scatter that renders a self-contained chart with no external scripts. You embed it with a standard <iframe>, so it works in any CMS, Markdown renderer or static site that allows iframes. Because the data lives on Vision Analysis, the chart stays current: when a model is re-benchmarked, every embed of it updates with no action on your side.
URL parameters
| Parameter | Values | What it does |
|---|---|---|
highlight | comma-separated model IDs | Models drawn in full colour with labels. e.g. yolonas-s,yolonas-m,yolonas-l. Highlighting a whole family draws its scaling curve. |
theme | light · dark · system | Colour scheme. system follows the viewer’s OS preference. Defaults to light. |
hw | hardware ID (optional) | Plot measured accuracy from this hardware instead of paper-reported mAP. |
rt | runtime ID (optional) | Runtime to read measured numbers from, e.g. tensorrt_fp32. Used together with hw. |
task | detection (default) | Which task pool to draw context dots from. |
Theming
Three modes ship out of the box. Light matches most documentation sites, dark matches dark-themed blogs, and system reads each visitor's prefers-color-scheme so a single embed adapts per reader. The background, grid, axes and tooltip all recolour together; model colours stay constant so a family is recognisable across themes.
Sizing & interactions
The snippet wraps the chart in a fixed 8:5 aspect-ratio box, so it fills the width of its container and computes its own height. There is no fixed pixel height to keep in sync, so no scrollbar appears whatever the column width. Hovering a dot shows the model name, parameter count and mAP; clicking a dot opens that model's page on Vision Analysis in a new tab.