AI driven keywording

I’ve decided to write about keywording this morning. For some time, I’ve used the IMS Keyworder website application which allows you to insert a few keywords into their web page and then choose which images match your particular one and then select the keywords from that appear in the various stock agencies against those images. I found that you can quite quickly identify photographs which match your image and then quickly choose the keywords that you think match the image in question. From there it’s simply a matter of copying the keywords into Lightroom and I make up appropriate titles and descriptions. I then use that completed image as the basis for copying keywords and descriptions across multiple similar files using the LR sync function. I can add a few particular keywords or change the descriptions as necessary to better describe the other files. My metadata gets inserted into the JPGs when I export them before upload to the stock agencies.

A couple of times I’ve tried the new AI-driven keywording systems where you can upload an image to a website and then obtain new keywords and descriptions for that image. However, this always takes me outside my main keywording flow which is that I keyword every processed image in Lightroom itself because that allows me to find images many years after I originally took them. With 140,000 images in my LR Catalog, that is a major requirement for me. Anything that requires me to export an image then import into an external system then copy the keywords back into my catalog seemed to be a step too far in my view.

I recently broke my wrist in a fall and so typing is relatively slow and difficult, and so I tried a couple of artificial intelligence driven keywording systems to see if they could help speed up the process. I went to Egypt in November and so I have many complex images to keyword. The first one I tried is from AI Stock Keywords and required an export of a file from Lightroom and then an import into their system, but it was quite easy to export a small file from my Lightroom catalog and upload that directly into the system. It is very simple – it just presents you with the title, description and keywords (which you can download in the file or in a zip file if you want). You can deselect keywords if you want. I did, because it includes silly keywords like photo, stock, royalty free etc. I just copied the keyword list and description and pasted into my LR catalog. The pricing is very competitive – just $2 for 1000 images!

The other system from Phototag.ai is more complex and gives you many more options in terms of the number of keywords, providing some context information such as the location of the file in question and how many characters you require in the title and descriptions. There is a Lightroom plug-in available which is easy to install and allows you to select a command within Lightroom when you have selected a photo, which must, behind the scenes, upload a thumbnail to the website and then generate and reload the keywords and the descriptions into the photograph and the LR catalog itself. It all works very smoothly and is much closer to my preferred way of working and avoids the need to export images, copy and paste keywords et cetera into the catalog within Lightroom. The pricing here is higher, but it is a much more complex system, at $14 for 1500 upload credits.

What I found was that the low-cost solution which required the export and upload of an image actually generated keywords and descriptions that better fitted the image in some of my test cases. For instance, the system correctly identified a particular temple in Egypt that I was keywording whereas the Phototag.ai system required that temple name to be inserted in the context area so that the AI system actually knew where the picture had been taken. This is not a great hardship, but does show how the world of AI keywording is moving forward.

This wasn’t the case on all photographs however and I must admit at the end that both of the systems actually generate quite reasonable keywords and descriptions. I still think I would add my own manual intervention to make sure that both the location and the correct details have been inserted into the file to allow buyers to eventually find and license that image.

I would have to do a lot more testing to finally rate one higher than the other! One of the systems is quite inexpensive and so it is very low cost to just use for each file and take the extra bit of trouble to remove spurious keywords. The Lightroom plugin approach is more expensive per image but still a relatively small amount compared to the time that these systems can save and the better keywords you will get, especially if English is not your native tongue.

I have affiliate links for both systems which I’ve inserted into the links above so if you would like to try these keywording systems, please make use of those links and I may receive a small commission as a result.

I’m not 100% sure that I know the answer to our keywording task. I think the IMS Keyworder system is still a great way of identifying similar images, but I must admit that the artificial intelligence driven systems are coming very close to what I would have chosen when I choose my own keywords from that system. I still like to add perhaps a few of my own keywords to make them slightly different from all of the other images in the agency databases but I will try each of the options as I go through these Egypt images and see which one I end up preferring.

UPDATE

After continuing to test various different images against these two systems, I’m concluding that the AI Stock Keywords system is better at identifying what is in the image, but the Phototag.ai system perhaps chooses more complex keywords for the image it sees. I tested two images of a riverside town in Egypt. AI Stock Keywords spotted the mosque and minarets in the image, which were not seen by Phototag.ai. Both recognized that it was an Egyptian town and chose various words describing the town and its surroundings. When I uploaded an image of our local pickleball courts (where I broke my wrist!), only AI Stock noticed that it was Pickleball, not tennis that was being played, but Phototag added Sports Complex, AI Stock did not:

Bottom line – they are both interesting. Phototag has much more control over what is going on, and is not particularly expensive. It probably produces better descriptions but they are still a bit flowery!

I’ve continued to play with both of these systems and finally, I think I have a conclusion. Each time I used AI Stock, I had to look at the keywords it produced one by one to remove what I think are redundant words such as “photo”, “stock” etc. Sometimes it provided complex multi-word keywords, sometimes it didn’t with no way to control that. If it was a simple file with not many obvious keywords, it tended to fill up the 50 slots with vague useless ones. I did have to export and do my copy/paste routine as well. Phototag allowed me to add a couple of keywords in LR (such as pickleball in the example above) and then just run their plugin to have it come back and fully fill in the results. I think I will normally suggest 30 or 35 as the required keyword number which reduces the chance of it returning less than useful ones. And you do have lots of control over what the results should be. So, even though it is bit more expensive, I’m going to choose Phototag.ai. A Winner!