Fashion Meets AI

Designing a Wedding Outfit with GPT-4

I wanted to write a post about a singularly interesting conversation I had with GPT-4 Vision recently. The impetus for this post was that I was genuinely excited to use GPT for a new kind of task for me.

The Task? I wanted GPT-4 to help me plan a wedding outfit.

Help I’m fashionably challenged and want GPT to help me plan a wedding outfit

The outfit had an interesting theme of “casual beach formal” , so I thought I creative AI should be able to help me out. With some simple prompts I asked to design an outfit for that theme and then I asked it to draw a image of that outfit on a mannequin with brown hair so that I could see it. I’m very visual.

A “casual beach formal” outfit as deemed by GPT-4. Oops I like it but that doesn’t look like me at all. Let’s try to tighten up this look customized for me.

I noticed the resulting image was quite beautiful, and as much as I make my resolutions every year, I don’t have a body like that. So I asked GPT for the mannequin in my dimensions, which I will not reveal here the resulting image striking similar to the first one.

Asking for several pant sizes larger and about 4-6 inches shorter. His brothers left but he stayed. Ok let’s try again

I tried harder to make the model my proportions, but the AI kept making him very thin.

No matter what I tried I kept letting the same Mr Hot-Man. For this image I specifically asked for love handles and to be fair I get where it was going with that around the mid section.

Getting bored and wanting something to happen, I asked it to do something more fantastical like draw an arm out of the mannequin’s head, it was happy to change the image for that. To be fair it didn’t do what I asked, but still.. it will rather draw robot arms than make me slightly fatter.

I asked for a third arm out of my head. That’s not what I got but hey it’s more willing to turn me into an android than to make a bit heavier so we can work from here maybe.

He is my son and I shall call him sinan v0.1 alpha. I even asked to make sinan v0.1 alpha stouter and shorter, and it still refused to do that for me.

Hey there’s that third arm! This is after I asked to make sinan v0.1 alpha stouter and shorter

It’s much easier to make a multi-arm mannequin than one slightly shorter and fatter I guess. As one more test I backed up a bit to the original outfit and I asked it to draw it for my Indian friend who wants to wear the same outfit. And here is what it came up with.

For me (on the left) and for my “Indian friend” on the right. Not much difference but the skin is slightly different for sure.

Honestly, at first I didn’t notice the difference but then after a few seconds I do see that the skin is slightly different on the right which I guess is fair. It’s just hard to get away from that chiseled face I guess, I get it.

Conclusion?

This is not a research study, nor did I do a lot of repetitive testing here. But it is a case study in my singular user experience with an AI to solve what I expect is not that rare of a request. As someone who has been working with Generative AI for over a decade, if I can’t get GPT to do what I want even with some minor prompt engineering, I wonder how long the average ChatGPT user will wait before rage-quitting on this exact scenario.

This is no way of criticism of OpenAI. Multi-turn multi-modal conversations is arguably the most challenging tasks being undertaken by commercial AI today. I’m only saying that any company who is building such AI experiences should remember to market not only the AI’s capabilities but also known limitations, no matter how small. It would at least be nice to know that other people see similar areas of improvement as I do.

Just a thought.