An Outfit You Could Wear in Sainsbury's and Get Only Admiration
Monday 18th December, 2023
Automated fashion design, Clothing photos, Corsets

You could wear this in Sainsbury's and get only admiration.

What it is is more output from ChatGPT4, a continuation of Saturday's dialogue. One of my corsets is the leather waistcoat I described there. It's great for posture, as it pulls the shoulders back. I'd seen similar items referred to as bodysuits, and I asked ChatGPT4 to show me one for a man. It did, this:

Having established that ChatGPT4 in conjunction with DALL-E could draw this kind of corset in isolation, I asked it for: "a man wearing such a corset as part of a complete outfit. The corset should be worn over a puffy shirt with long gathered leg-of-mutton sleeves. Its wearer should also be wearing a voluminous velvet skirt, of luxurious smooth sheened velvet adorned with electric blue lightning flashes. The corset should be tight soft black leather, with one lightening flash centrally placed. The puffy shirt should have a colour or colours which harmonise with these. The man should be 6 foot 4 tall with a 25 inch waist and blond hair." The result was this:

What's going on here? Current "AI" drawing programs are not terribly good at relationships. They don't really understand words such as "on" and "at". Consequently, they tend to "leak" attributes and components, losing them or putting them in the wrong place or with the wrong object. They are also not good with sequences of repeated items such as buttons and the peg-and-eye fasteners on the front of a corset. When I asked ChatGPT4 to draw "a table. On it is a red brick. On the red brick, is a green brick. To the left of the red brick is a blue brick. To the left of the blue brick is a violet brick. On the violet brick is a carmine pyramid", it produced this:

The colours are there, and so are the tabletop and the pyramidality. But the bricks are certainly not arranged in the way I specified. And something similar has happened with my design. There is an attribute of leatherness, but it has leaked into the shirt. The front of the corset appears to have some kind of silver fastener, but it also has lacing, which would normally be at the back. I don't see any flashes on the skirt; and that on the corset is not central. Nevertheless, it's a striking outfit, and one I'd happily try. And "AI" programs will improve — there are some very clever people trying to overcome their limitations.

Going back to dress, ChatGPT4's picture didn't show the man's feet. I asked for one that did, but this drifted the design considerably. The puffiness of the leg-of-mutton sleeves is leaking into the legs, trousers are appearing, and on the back of the third figure, a flash continues from corset to skirt:

I then asked for a real-world setting: the shelves in Sainsburys, shopping for a jar of Sun-Pat peanut butter and a tin of baked beans. Would the context tone down the design?

It does tone down the exuberant volume, but the guy looks too heroic. His hair is unmessed, but still, you'd expect to see him conquering dragons, not chasing discounts. So: "Make the man be putting the peanut butter jar in a Sainsburys shopping trolley. Keep his blond hair, and keep his outfit exactly as it is. But make the face a bit rounder, with scholarly black-framed round glasses. The hair should be a little bit messy, as at the end of a working day."

I then tried to restore the costume to the original topic: "Give the man a long corset as in the previous image, but keep the rest of the attire."

But: "The man in the image above that has a calmer, rounder, less morose face."

I'm eagerly awaiting the time when we understand the mathematical space inhabited by these programs well enough to ask for conceptual search-and-replace: change the face, or change the skirt, but leave all else the same. That will come. In the meantime, one can still generate some stunning designs.