diff --git i/app/helpers/application_helper.rb w/app/helpers/application_helper.rb
index 8566ce026..18a56ebae 100644
--- i/app/helpers/application_helper.rb
+++ w/app/helpers/application_helper.rb
@@ -960,6 +960,7 @@ module ApplicationHelper
 
     text = text.dup
     macros = catch_macros(text)
+    contains_macros = macros.any?
 
     if options[:formatting] == false
       text = h(text)
@@ -978,6 +979,7 @@ module ApplicationHelper
       end
     end
     parse_sections(text, project, obj, attr, only_path, options)
+    fix_paragraphs_with_block_level_content(text) if contains_macros
     parse_headings(text, project, obj, attr, only_path, options)
 
     if @parsed_headings.any?
@@ -1498,6 +1500,76 @@ module ApplicationHelper
     end
   end
 
+  unless const_defined?(:PARAGRAPH_TAG_RE)
+    PARAGRAPH_TAG_RE = /<p(\s[^>]*)?>/i
+    PARAGRAPH_BOUNDARY_RE = %r{</?p(?:\s[^>]*)?>}i
+    BLOCK_LEVEL_CONTENT_RE =
+      %r{<(?:address|blockquote|div|dl|fieldset|figure|form|h[1-6]|hr|ol|p|pre|section|table|ul)[\s/>]}i
+  end
+
+  # Macros are replaced by a placeholder before the text is formatted and are
+  # executed afterwards, so a macro standing on its own line has its output
+  # injected into the paragraph the formatter built for that placeholder.
+  # Macros returning block-level content ({{include}}, {{collapse}} and macros
+  # provided by plugins) therefore produce block-level elements inside a <p>,
+  # which is invalid: browsers close the paragraph before the block element and
+  # turn the then orphaned </p> into a second, empty paragraph. Both show up as
+  # unexpected blank space around the macro output.
+  # Such paragraphs are turned into divs to keep the markup valid.
+  def fix_paragraphs_with_block_level_content(text)
+    return text unless text.include?('</p>')
+
+    result = +''
+    scanner = StringScanner.new(text)
+    until scanner.eos?
+      preceding = scanner.scan_until(PARAGRAPH_TAG_RE)
+      if preceding.nil?
+        result << scanner.rest
+        break
+      end
+
+      opening_tag = scanner.matched
+      attributes = scanner[1]
+      result << preceding[0...-opening_tag.length]
+      content = scan_paragraph_content(scanner)
+
+      if content.nil?
+        # Unbalanced paragraph, leave the remaining text untouched
+        result << opening_tag << scanner.rest
+        break
+      elsif content.match?(BLOCK_LEVEL_CONTENT_RE)
+        result << "<div#{attributes}>" << content << '</div>'
+      else
+        result << opening_tag << content << '</p>'
+      end
+    end
+    text.replace(result)
+  end
+
+  # Returns the content of the paragraph the scanner has just entered and moves
+  # the scanner past the matching closing tag. Nested paragraphs are taken into
+  # account, so that a paragraph belonging to injected macro output does not end
+  # the enclosing one. Returns nil if the paragraph is never closed.
+  def scan_paragraph_content(scanner)
+    content = +''
+    depth = 1
+    until scanner.eos?
+      chunk = scanner.scan_until(PARAGRAPH_BOUNDARY_RE)
+      return nil if chunk.nil?
+
+      tag = scanner.matched
+      content << chunk[0...-tag.length]
+      if tag.start_with?('</')
+        depth -= 1
+        return content if depth.zero?
+      else
+        depth += 1
+      end
+      content << tag
+    end
+    nil
+  end
+
   TOC_RE = /<p>\{\{((<|&lt;)|(>|&gt;))?toc\}\}<\/p>/i unless const_defined?(:TOC_RE)
 
   # Renders the TOC with given headings