Ongoing Copilot experimentation

 I continue to explore the use of my university’s Microsoft Copilot license as a generative AI diffusion model in the area of costume design rendering. 

For this iteration, I attempted to train Copilot on my own renderings by instructing it to look at images from posts I made to a long-abandoned blog, A Sketch A Day. 

The concept behind the blog was to encourage the cohort of participating costume designers to share images from their sketching process, both preliminary sketches and final renderings. There were several participating designers, and the goal was to share at least one sketch a day, from at least one of the designers, although several of us could post on a given day if we had material to share.

At the time, it appeared that I might be designing more frequently for the theater in residence on my campus, and I thought I would be generating more design sketches than I had since my undergrad days. I lacked confidence in my sketching ability and participation in the blog gave me motivation and accountability to maintain a regular sketching practice. 

But regarding this post, I thought I could instruct Microsoft Copilot to pull sketches from the blog and use them to inform the style of the costume design rendering I prompted it to produce. I hoped it would generate a version of the Elphaba series that looked like I could have drawn it.

In this case, my mistake was presuming that my university’s Copilot license would allow it to pull images from the blog. But Copilot metaphorically clutched its pearls—how does it know that those images are of sketches I have permission to use? 

Never mind that none of the training data was used with permission or consent but I guess Microsoft is putting toothpaste back in the tube now.

So it told me:

I can't directly access or replicate the specific style from the blog posts, but I can create a new sketch inspired by the Art Nouveau style, similar to Alphonse Mucha's work. [1] Let's try again with your detailed description:

• Full figure of a muscular, sturdy green-skinned witch
• Dress covering the shoulders with a high collar, trimmed with leaves and vines
• Hair in long braids of green, gold, and brown
• Rustic, angular witch hat

I'll generate the updated sketch now. Please hold on for a moment!

I apparently chose to interact with Copilot at an exceptionally popular time (Sunday afternoon). It took the model FIVE HOURS to generate the images. 

But this round is actually much better, although she’s still willowy & ten heads tall. Here’s what it gave me:

'

The closest to what I was going for 

(I had those boots in the 90s!)

Another decent iteration despite the creepy smile

Not bad although why’s she wearing a transparent skirt?

Ah yes, in case I was lying when I said 

the earlier naked wasp-waisted versions were unacceptable

I don’t know. At this point I guess I can’t imagine using some kind of generative AI model in a costume design application, unless something drastically changes about how production design occurs. I’m not seeing a way to exert the kind of fine-grained control over the output that a costume designer so often requires.

Perhaps it’s helpful in teaching complete novice students learning how performance costume design works, to keep them from getting too hung up on the quality of their own sketching? Although collage would probably be easier and more controllable. I don’t know.

I’m intrigued that Copilot only began to approach following my prompts after I told it its work was terrible. I think it’s counterproductive to ascribe emotions or sensitivity to a machine learning model. At the same time, the conversational chat interface creates a disturbing sense of personality such that it’s easy to imagine Copilot as an eager intern desperate to please.

Do I have to talk to it like a drill sergeant to get it to obey? 

Joking, but not. I mean, I have to practically yell at Alexa, & the processing bias of home assistants against non-masculine voices is a documented thing.

[1] Mucha was my original style reference three prompts ago.