I have always considered myself an artist in one form or another. Most of which I have never been particularly good at from a technical or skill standpoint. One of my weakest areas must be drawing. Conceptualizing what I have visualized, pulling it out of my brain and onto the medium in front of me, has always been a frustrating act of futility. Yet when things like Dall-E and Midjourney came along, I thought maybe the shortcut of my dreams has finally arrived.
Or maybe not…
I haven't played much with Midjourney yet, but Dall-E I have spent more time with. I still use ChatGPT extensively in writing site copy for my clients. I always have it open in a tab, leaving plenty of time to play around with Dall-E. Yes, I am that guy drawing pictures of their pets, slowing the works down for everyone else.
Getting Dall-E to listen can be rather challenging. The same point can be made repeatedly, and it will not listen. I have some suspicions on what caused the issue with the images below, and it may be obvious right away to some, but still certainly an issue. I started out by giving it some real images of my Shih Tsu, Huey, so it would have something to source from. I did not give it an image of myself but gave it some basic guidance on my appearance. Here is the prompt use, and the resulting image.
Prompt:
Draw a picture of me, I have lots of muscles and guns, long dark hair, and a beard, alongside Huey and we are having a blast down on the Miami strip
Not bad, overly cartoonish, but I asked for something a little bit out there. Dall-E can really struggle with photo realism, but that does not seem to be the point of it overall, so I will not be knocking it for that. I would never be so reckless with a firearm in public, or around my pup, so I decided to try something that would match our vibe a little more.
Prompt:
Let's move us to a dive bar on the beach and give me a Marilyn Manson shirt. Both Huey and I should be drinking whiskey.
I am not sure what a "bansy" is… A weird mix of "band" and "Banksy"? We know all these generative art applications struggle with lettering, and it did good with the text I did provide so I am willing to overlook it. I do not have any tattoos, however, so that is going to be my first edit. Similar prompt as before just a minor tweaking.
Prompt:
Let's move us to a dive bar on the beach and give me a Marilyn Manson shirt. Both Huey and I should be drinking whiskey. No tattoos at all ever.
This is one of my favorite images it has produced. but I still clearly have tattoos. Sometimes I find just asking ChatGPT why it generated content will provide valid reasoning, and is a fantastic way get more efficient responses.
Prompt:
Why are there still tattoos?
It gave me no explanation, and simply created this image. This one is not bad, though I do not think it looks much like Marilyn Manson and it still features tattoos. I think Manson being the cause. Any training it may have of what Marilyn Manson looks like will almost certainly feature his arms, adorned with many tattoos. Some of which are like the ones seen here. Or perhaps it is trained on images of people wearing band shirts who had tattoos themselves. This, to me at least, seems like it is the most likely reason for its struggle with this task. At the very least, it does provide images in WebP so I don't have to bother with converting them, which is nice I guess.