The Hello Kitty Prompt Project - Part 1
Santiago Melo
13 years of experience structuring and developing digital transformation business models that focus on the human element and are supported by performance marketing and growth modeling tools.
As part of the work I am doing to understand generative AI I wanted to play a little with it. While there are many AI Image Generators most have limits on the number of images and I don't yet have a project that would warrant paying for MidJourney or Leonardo AI. Microsoft Designer and it's integration of DALL.E 3 which allows unlimited creation of images is a fun way to place with image generation (https://designer.microsoft.com/image-creator ).
Amongst the different tools provided by Microsoft are pre-generated prompts. These are a great way to play with the system, but before we play with them as you likely guessed this article is about what happens when we ask the system to draw Hello Kitty (https://en.wikipedia.org/wiki/Hello_Kitty ), the classic Sanrio character, and then place it on a variety of prompts.
The first prompt is to ask what it thinks Hello Kitty is:
Not bad, this is a very accurate "drawing" of the character. It is interesting how it is trying to build a scene around Hello Kitty without any prompt. In one case it goes more for pastel colors while the other uses more saturated tones. I am baffled as to why it decided to add multiple mini-Hello Kitties and other strange animals. Lets see if we can get the character without any background.
This was interesting... Just by adding the blue background we have completely changed the style leaving behind the cartoon drawing of the first prompt and now going to a 3D rendition of Hello Kitty. While in the first try we had two very similar scenes, these are two different interpretations. And truly chef's kiss to the detail of a smaller Hello Kitty in the yellow sweater.
As we give the prompt more direction it tightens the possible interpretations and is able to better predict what we want to see. As we change the background color interesting things happen as well. Notice in the picture below how no longer we are seeing the character from a turning profile, but the change of color also seems to guide the intention that we want about how we need to see the character.
As we look at the two pictures above, it is important to question how the system predicts and interprets what the output should be based on a parameter change.
Green seems to have produced a very similar result to pink, but look at the floor on the first image, maybe an indication that we need to build a house for Hello Kitty. We should start by the floor.
The second image provides a better interpretation of the prompt, but the second image is a better predictor of the outcome we wanted. Let's try to see if there is something we can add to our text that can replicate this outcome.
领英推荐
So now Hello Kitty has stopped existing in a color void and is actually in a place with walls. Where is the light coming from? Maybe we need to add a window to the wall to explain this or could we play with the light?
A couple of lines above I used the term prediction. It is important to clarify that generative AI is not predicting, it is truly generating new content. The prompts we are writing guide that production and it can feel like it is predicting the outcome we want to see. Here are some links to help clarify this difference:
Well now Hello Kitty has a window that looks outside, although the first image is giving me door vibes, maybe the system is telling us that we need to give it a door.
This took several tries, in most cases there was no window or no door. Finally the second image here contains a sliver of a window, but that could be the aspect ratio.
Changing the aspect ratio didn't seem to do much on the size of the image provided but it does seem to have affected the image generated. In other tools you would actually get a different aspect ratio for the image. But the house is looking very empty, what happens when we add some furniture to it. Currently our prompt looks like this: Hello Kitty standing by itself on a wooden floor. Pink plain color wall with a window. The wall has a white door. There is a variety of furniture in the house.
Since the software is interpreting our prompt, this becomes the most important part. Understanding how much we need it to change in order to show what we want., keeping in mind that something might get lost.
I think leaving Hello Kitty sitting in a couch is a nice way to end this first part. Our prompt currently reads like this: Hello Kitty sitting on a couch by itself on a wooden floor. Pink plain color wall with a window. The wall has a white door. There is a small dresser beside the couch with a plant on top. The final images produced after some experimentation with the above prompt are this.
Although what is happening here:
What an excellent outcome, and thank you for the detailed prompts! I haven't used Microsoft Designer yet, but after reading your article, I will try it.