Who Let the Dogs Out??
Is digital always better? In this blog post I take a look at how Craiyon.com, a text-to-picture tool, processes the common saying “Who let the dogs out?”
I decided to draw the saying “Who let the dogs out?” and was surprised by the output. I assumed that the program would just aggregate a mixture of relevant photos but instead it appeared to mesh a combination of images, which resulted in a panel of shots that do not appear to be quite right. It is clear from the pictures that the animals are supposed to be dogs, but they are slightly warped. The AI program did get the “out” portion of the saying though, because the images are all taken outside.
Though the end product was not flawless, I was impressed that the program was able to convert words into fairly realistic images related to the topic.
I think the program also has a hard time understanding a common phrase – it takes the words as individuals unless separated by really clear conjunctions. I would argue that the AI really missed the “who” portion of the prompt! I feel like through my experimentation, the tool was able to make a good portrait of 50% of the prompt and then ignored the other half. Its funny that the tool can make dog’s faces better than human faces!
Thank you for the post, I tried two different phrases 1) be like a dog with a bone 2) Underdog and both gave very surprising results, for underdog it gave a dog cartoon character dressed in a super hero suit and for be like a dog with a bone the result was few distorted faces of dogs (majority of the dog faces were better than human faces) with a bone. It is very clear AI has a hard time interpreting the idioms and sentences.
This is so funny. The AI got so close!! My favorite is the bottom middle dog, which truly looks like a collage of other dog images (which is essentially what the AI is doing). Also what an interesting thing that the AI is better at compiling animal features than human ones! Or maybe we’re just better at discerning human features rather than dog features. Interesting example!
Amy, thank you for the post. I assume AI would have a more difficult time analyzing when there are more words (and when they become a phrase or sentences). Interesting to see some dogs have a leash and some do not. Also fun to see how it understood ‘out’ as green fields. I guess some random objects (leaves, balloons, and toys) with dogs show Craiyon’s effort to do a better job 😉
Thank you for the post, Amy! I think Craiyon’s images of the dogs (though slightly warped!) are absolutely adorable. It’s a stark contrast to the scary images it produces of humans – perhaps it will take Craiyon several more iterations to display humans properly. A part of me was hopeful that Craiyon may interpret your phrase with some nod to the Baha Men song!
I agree with the commenters that Craiyon has a difficult time with certain phrases – I tested several. Some phrases that are extremely common parts of our culture (“Star-spangled banner” , “The right to bear arms” , “Et tu, Brute?”) yielded very clear, expected results whereas others (“Who let the dogs out” or “There’s no place like home”) perhaps had too broad of an interpretation.
Hi Amy, I love the topic you chose! Even though the results from the tool were not perfect, I still can see some patterns emerging from all the images (green grass, dogs with toys, cute faces). I think the tool is smart enough to retrieve a series of characteristics that people usually appreciate and look for when they search for certain results.