It’s the week before Thanksgiving, and I’ve overheard my students making plans for “Friendsgiving”, and frequently talking about vegan options. So I wanted to challenge myself (and my AI tools) with a difficult assignment. Could I get a decent art directed image that conveyed the feeling of “Vegan Friendsgiving”. Originally, I also wanted to reference the famous Norman Rockwell painting for composition, but Dalle-3 shut me down on the use of his name, so I let it go.

I wanted to make sure the image include a diverse group of young adults and vegan food options, but I didn’t realize how difficult this would be. To be fair, large groups of people and a complex table full of food is a tall order. I tried a lot of different styles from photography to illustration, and he goal was to make everything a 16:9 aspect ratio, but some tools weren’t that flexible. Almost every tool failed on one criteria or another. To keep things moving, I allowed a maximum of 10 minutes on Dalle-3, Midjourney, Adobe Firefly, Ideogram, and Leonardo. From a pile of images, I chose the top four. Brace yourself, here are the results.

#1 Dalle-3 with help from ChatGPT

Chat GPT wrote some careful prompts which produced some intricate images. The hero image at the top of the story (with it’s funny witch hats) was from this batch, but these illustrations with the crowded folk art quality are really interesting. Response time was really slow.

#2 Midjourney

No surprises, Midjourney really delivered here. Both the photos and the illustrations captured the look and feel that I was seeking. Almost every prompt had at least one good image and I found results I liked really quickly.

#3 Adobe Firefly

You can’t be serious Adobe. While the illustration in the first slot is pretty good, the photos (those mouths!) are pretty atrocious considering your training data is all from Adobe stock. This should be easy for you. I tried a lot of different styles, but the results kept getting worse.

#4 Ideogram

Ideogram really struggled here, and the more I prompted, the uglier the results got. In desperation, I tried a text-only option, and that turned out pretty well so I included it. Overall, not ready for prime time.

#5 Leonardo

While some of these are a little rough and the facial features could be cleaned up, they all have a certain charm and whimsey that I am enjoying. Leonardo gets extra credit for a easy to use interface that delivered great images really quickly.

Conclusion

I knew that this was going to be a challenging assignment, and I’m not disappointed with the results. While some of these images are not quite there yet, it was a valuable exercise and provides a great snapshot of the current state of the art across the board. Given more time and perhaps some more thoughtful prompting I probably could have coaxed out some better results. However, to really get the results I want, I would definitely need to resort to in-painting, using Photoshop in a traditional manner, or the new generative fill. I’d probably also need to run them through an upscaler so I had enough pixels to fiddle with. We’re certainly not at the point were we could clunt on just one tool for all our image generation needs.

‍