We built an AI system to auto-tag all your inspirations
In our earlier post, we’ve talked about the pain of manually tagging our image libraries. In order to take full advantage of the tag systems, you have to manually assign tags for each image. It could take a lot of time to get the tags ready. If you already have thousands of images, this would become an impossible mission.
So we built features like random sorting and InspireMe!, so that you can find related contents without any manual settings.
Those features worked well when you want to get some random inspirations. However, if you want to be more specific, a tagging system still seems to be inevitable.
But do we really need to tag all the images manually?
We hope not. That why we built an AI engine fine-tuned for design contents. The InspireMe! feature was our first step to bring it to life. We’ve been training it intensively since then. And now, we think it’s time to push it to a new level.
Here is a sneak peak of what’s coming in our next version:
As shown in the video, when an image is selected, it will be analyzed immediately. A group of suggested tags will be listed on the properties bar. You can simply click to commit the tags, or click on the “x” to remove them. All you need to do is a few clicks, and the images will be properly tagged (you can still type in manually if you want).
“Hmm.. but it still takes lots of clicks if I have thousands of images?”
As a matter of fact, it is not necessary to do it for all your images. Till now, we’ve trained our engine with over 30 different design categories, and thousands of common objects. It is impossible to make sure all the predictions are correct, but with that much information, we are able to calculate the differences between each image. When you make some manual corrections to some of the images, the engine will be able to learn from it and adjust the predictions to their nearest neighbors. The more tweaks you make, the more precise the predictions will be for the rest of the images.
Even if no manual corrections are made, you will still be able to search and filter contents with the predicted tags. They might not be very precise, but again, wouldn’t it be fun to get some surprising results when searching?
By the way, everything happens locally, nothing will be collected or uploaded during the prediction.
How good is the prediction?
As said, we’ve trained the engine with over 30 design categories, plus thousands of common objects. However, the user data could be very different and complicated. We still need time to keep improving the engine, but we hope this could be our first big step toward our final vision.
Enough talking, show me the product
Well, we are still in the final stages of testing and finalizing the auto-tagging system. Be sure to follow us here in Medium or on Twitter, we will post more updates very soon.
update: the version 2.1 is out, check this article for details of the new features.
