A few weeks ago I blogged about content based image retrieval and listed a bunch of resources to learn more including several items from a company called LTU Technologies. At that time a content based image retrieval demo using LTU Technologies was unavailable. Good news, it's back online.
A visit to LTU Visual Search allows you to search a database containing 65,000 royalty-free images from Corbis, a stock photography service. You can search keywords associated with each image (I believe from the Corbis controlled indexing vocabulary) or browse random images. Next to each thumbnail is a button labeled "Find Similar Images." Clicking the button allows you to find similar images based on both the keywords AND the visual characteristics of the image. More about all of this in my original post and this help page.
Note: The full Corbis search interface (sans the LTU technology) offers numerous ways to focus and limit a search that you don't see in most of the large open web image databases. Why don't services like Flickr also allow users (optional, of course) to include fielded metadata along with keyword tags? If extra data was included with each image record (in addition to keyword tags), it could potentially make searching more powerful and precise for all searchers.