Oh I see. Indeed, I couldn't find any documentation on that and only looked at a couple of examples, my bad. Can you please tell me where to find those? I could only find this one which however doesn't cover all cases (ie. is there one shortcut for image path as well?).

You don't say why this would be useful in your case -- could you elaborate?

I'm writing a filter to handle wrapfigure in LaTeX (and maybe other formats in the future, it would be cool if it could target HTML and DOCX/ODT as well). This is the syntax which I have in mind:

![Image caption](/path/to/image.png){.wrap width=3}

As I've previously pointed out the manual defines a specific behaviour for width and height attributes and states that when no unit is passed it's assumed to be pixels. However, in order to convert pixels to inches (without introducing a misleading behaviour not matched by the manual) for LaTeX output, the filter needs to know the DPI value, which can be overridden from the command line: knowing such value would allow me to write a function to handle different units passed (or not passed) as part of the value to the width attribute, complying with this section of the manual. Still, given the complexity of the task I think it would be a better idea to let Pandoc handle this once for all, converting accordingly to the output format:

Dimensions are converted to inches for output in page-based formats like LaTeX. Dimensions are converted to pixels for output in HTML-like formats.

Hope I made myself clear.

I'm reluctant to do that unless we add Underline as a proper INline constructor

I agree with you.

More generally speaking, can we rely on classes and attributes as part of the sematics? If not, I think I have to redesign the syntax for this particular filter.