Tattoo Time (Part II)

Getting LLMs to support my tattoo adventure
LLMs
prompting
tattoos
image generation
computer vision
image recognition
Author

Leesa

Published

March 15, 2024

image.png

Getting Creative

In my last blog, I had ChatGPT and CoPilot help me design a tattoo based on my yoga retreat experience. However, the designs that they produced weren’t quite what I had in mind. I asked around and my friends gave me a couple of good ideas for a tattoo, so I thought it would be interesting to see if ChatGPT could help me to design a tattoo based on specific instructions.

Initial Prompt

I gave ChatGPT the following prompt: You are a very talented tattoo artist. You have a client who has come in with a very specific design request: a tattoo of a young, black woman with short hair standing in a field of low flowers. The flowers are outlined in fine detail, but the woman is a shadowy silhouette. There is a large moon behind her backlighting the scene. Her arms are out, hovering over the flowers at her feet. The flowers only come up to her mid-calves. The tattoo should be in black and white.”

It responded with the following image: image.png

Now, this is not at all what I had in mind.

I double-checked the underlying prompt that ChatGPT ran to determine if there was a disconnect between my instructions and the prompt it used. It ran: A striking black and white tattoo design featuring a young, black woman with short hair. She is depicted as a shadowy silhouette, standing in the midst of a field of low flowers that are outlined with fine detail. The scene is backlight by a large moon, casting an ethereal glow and creating a contrast that highlights the intricacy of the flowers while leaving the woman in silhouette. Her arms are outstretched, hovering gracefully over the flowers at her feet, which only reach up to her mid-calves. The composition balances the simplicity of the silhouette with the complexity of the floral details, creating a harmonious and captivating image.

This is very close to what I input so I figured the prompt wasn’t the problem. Instead, I wondered if the style of this image might have something to do with what we discussed in class about image generation systems having a predisposition toward fantasy material.

!

Providing Context

To solve this problem, I thought that providing examples might be helpful. This could show ChatGPT that I wanted a tattoo that was a bit more realistic.

I prompted it: This is pretty good, but not quite what the client wants. She has provided some images for reference to help guide the process. Please use these to inform your design. Remember: the woman is a black woman with short hair, and she is in silhouette. The image should be very simple

I included the following images: image.pngimage.pngimage.pngimage.pngimage.pngimage.png

ChatGPT sent back: image.png

This was much better! It looks like the reference photos really helped.

Final Touches

While I liked the most recent version, I wanted to see if we could get it closer to the image that I had in my mind.

I sent ChatGPT this image that I actually drew myself: IMG_5195.jpg

It sent back: image.png

I love it!!

This was a very useful prompting exercise to help me utilize some of the specific tools that we discussed regarding prompt engineering, including giving ChatGPT a role, analyzing the underlying prompt that ChatGPT ran, and using examples to help guide its behavior