Google Teaches Computers How To Pick Out Better YouTube Thumbnails

As we’ve noted before, the quality of a thumbnail image can make or break the success of the video it’s attached to. Creating thumbnails is an intricate art, but it’s not so intricate that a computer can’t do it properly. In a blog post, Google’s research team explained how it is using machine learning to automatically generate proper thumbnails.

YouTube has always been able to automatically generate thumbnails for videos uploaded to the site, but those thumbnails were simply fetched from a point toward the center of the video. Therefore, many videos have received blurry, out-of-focus, and irrelevant thumbnails.

The new “thumbnailer” is much more accurate. It uses the custom thumbnails created by YouTube video uploaders as its database, and it employs the data it gleans from those custom thumbnails to pick out the most thumbnail-friendly points within new videos. It then converts those stills to a thumbnail-friendly resolution. In the image above, you can see how the new thumbnailer’s chosen images compare to the ones generated by Google’s previous system.

Viewers have noticed the improvement. Google’s post cites a study in which human respondents preferred the new thumbnails 65% of the time. That’s not perfect, but for an automatic service that runs adjacent to the custom thumbnail uploader, it’s pretty good.