Critical Web Fonts
The history of web font loading has included many different iterations:
- Doing nothing: including a
@font-faceCSS block and using it in your CSS without qualification. I consider this to be an anti-pattern. It introduces a Flash of Invisible Text (FOIT) in some browsers and worse, creates a single point of failure in browsers that do not have a font loading timeout (WebKit).
- Data URIs for the Flash of Unstyled Text (FOUT): Loading a CSS stylesheet asynchronously (or using an AJAX Request) with the web fonts embedded as Data URIs (and putting it into localStorage for repeat views). This approach is deprecated because it can introduce a short Flash of Invisible Text (FOIT) on some older lower powered devices.
- A Scoped Class for the Flash of Unstyled Text (FOUT): Using the CSS Font Loading API (or a polyfill to do the same, like FontFaceOnload or fontfaceobserver) to add a scoping class protecting our content from our web fonts before they load (also documented on the Filament Group Lab). This is what I’d consider to be a bare minimum best practice, or Intro to Font Loading 101.
- Two Scoped Classes for the Flash of Faux Text (FOFT): This method complicates things a little bit more and uses two different stages of scoped classes. The first stage loads only the Roman font, then any variations of that are loaded in the second stage: Bold, Italic, Bold Italic. It’s great for slower connections in that it puts the biggest amount of reflow later in your page load waterfall to make it less noticeable and less impactful to end users. I’d classify this approach as Intermediate Font Loading 201.
The Next Iteration: Critical FOFT
This method builds on the Flash of Faux Text (FOFT) using a two stages loading process (sounds like, but is not actually rocket science), but instead of a full Roman web font in the first stage, it loads a small subset of the Roman web font, in this case with only the upper case and lower case alphabetic characters. You could optionally include numbers here as well.
I have implemented this technique live on my web site. Here I use four webfonts: Lato Roman, Lato Bold, Lato Italic, and Lato Bold Italic.
- The original Lato Roman is 25KB in WOFF2 format.
- Lato Roman with only the
A-Za-zglyphs is only 9KB in WOFF2 format (36% of the original)
The shrinks the first stage quite significantly.
The yellow frame highlights when the roman web font (used for most body content) has loaded and rendered. Note that all of these film strips were captured on zachleat.com using Chrome’s “Regular 3G” network throttling feature.
Scoped class for FOUT
Two scoped classes for FOFT
Two scoped classes for Critical FOFT
We’re essentially lengthening the load complete time at the end of the waterfall to reduce the jarring reflow shown to the user. Once
font-size-adjust is implemented in more than just Firefox, the need for this technique will be reduced, if not rendering it unnecessary. But until then, I consider this approach essential to reducing interruption to the user’s reading flow.
Enable 3G throttling in your favorite web browser (some of you may not have to enable anything to have a slow connection) and watch the page render. The web fonts do feel faster, even if the total time spent loading web fonts is greater.
The code used to implement this approach can be viewed on GitHub. It’s using FontFaceOnload for the font loading polyfill.
If you’re interested, I’ve also implemented the same approach using promises and Bram Stein’s FontFaceObserver.
For repeat views, you can continue using the same mechanism you were using before. I use Bram Stein’s
sessionStorage trick so that I don’t have to do anything server-side. You can see the implementation in code on GitHub in initial.js and fonts.js (also lines 15 and 37).
Ideas for Improvement
Some font loading services already provide mechanisms to subset and modify your web fonts dynamically. Dynamic Subsetting generates a subset of a font on the fly using only the glyphs used on a page. Dynamic Augmentation add glyphs to an already loaded font on the fly, sort of like font streaming.
You could use extend this approach using something similar to the Critical CSS workflow and scan the page to find out what actual characters are used in a fixed viewport (
In fact, you could wire up something similar to this using the amazing Plumin.js but unfortunately the library is too large (~
400KB minimized) for this use case. For now, I’ll just stick with a simple
9KB baseline WOFF2 that gets replaced with a
Live Font Interpolation on the Web, an article written for A List Apart by Andrew Johnson is also worth mentioning. Keep an eye on this approach—it would allow us to dynamically generate weights and styles from font masters, hopefully saving bytes on the wire for designs using a wide variety of weights and styles. I’d love to see this web-standardized.