@Edent a complete micro-optimisation but putting repetitive attributes first and ones with higher entropy values after might be better for compression?

e.g. <img sizes=“” srcset=“”>

Sizes might be the same across many images but srcset will differ.

@andydavies not to mention that in-viewport images _should_ just be served in a progressive format, so their first bytes can arrive milliseconds after document TTFB, this should mean no need for LQIP on those.

Outside the viewport, LQIPs could just be low-priority requests that are excluded from the preparser, like `loading=lazy`.

@andydavies well if we're talking preferences... I'd prefer it implemented directly into an image format like JPEG-XL and we extend the `loading` attribute to allow `loading=preview` (or `loading=metadata`, like `<video preload>`) only request the first however-many bytes needed for LQIP.

A range-request could then be sent for the rest when needed.

Then we can have image transformation layers like IMGIX and Cloudinary serve our LQIPs automatically from source images.

@sophie I feel like because the defaults were so damn ugly and there were limited options for design back in the day, people got used to just removing the underline.

If browsers had more aesthetic defaults (skip ink + wider offset + line colour being a 50% transparent version of the link's text colour, for example), maybe people would be less averse to them and not just default to removing them completely.