The wiki at the google-diff-match-patch project shares some ideas. From
http://code.google.com/p/google-diff-match-patch/wiki/Plaintext :
One method is to strip the tags from the HTML using a simple regex or node-walker. Then diff the HTML content against the text content. Don't perform any diff cleanups. This diff enables one to map character positions from one version to the other (see the diff_xIndex function). After this, one can apply all the patches one wants against the plain text, then safely map the changes back to the HTML. The catch with this technique is that although text may be freely edited, HTML tags are immutable.
Another method is to walk the HTML and replace every opening and closing tag with a Unicode character. Check the Unicode spec for a range that is not in use. During the process, create a hash table of Unicode characters to the original tags. The result is a block of text which can be patched without fear of inserting text inside a tag or breaking the syntax of a tag. One just has to be careful when reconverting the content back to HTML that no closing tags are lost.
I have a hunch that the 2nd idea, map-HTML-tags-to-Unicode-placeholders, might work better than one would otherwise guess... especially if your HTML tags are from some reduced set, and if you can perform a little open/close touchup when displaying interleaved (strikethrough/underlined) diff markup.
Another method that might work with simple styling would be remove the HTML tags, but remember the character-indexes affected. For example, "positions 8-15 are bolded". Then, perform a plaintext diff. Finally, using the diff_xIndex position-mapping idea from the wiki's first method, intelligently re-insert HTML tags to reapply stylings to the ranges surviving/added. (That is, if old positions 8-13 survived, but moved to 20-25, insert the B tags around there.)