The 2023 Microsoft Work Pattern Index report, “Will AI Repair Work?,” landed simply weeks after the launch of Bing Picture Creator, a brand new design software powered by a complicated model of DALL-E. When it got here to creating artwork for the report, the group noticed a singular alternative to discover how a human artist would incorporate it into their course of. Would it not assist generate concepts? What prompts can be simplest? And the way would utilizing AI as a co-collaborator influence the ultimate product?

The artist for the job: Los Angeles-based Jon Han. Han makes a speciality of creating thrilling futuristic worlds with out the standard dystopian tropes. As a result of the report knowledge confirmed that most individuals are longing for AI to assist ease the burden of labor, his type felt like precisely the fitting strategy to deliver the findings of the WTI report back to life.

Jon began by feeding prompts into Bing Picture Creator, testing how he might collaborate with it. Utilizing AI impressed a extra iterative strategy than what’s typical, as a result of taking part in round with it was fast and simple. “The method for this mission was an experimental one,” he says. “Simply making an attempt to see what it’s able to.”

Right here, Han takes us inside that course of.

Jon Han: In my regular course of, I learn a artistic temporary, take into consideration the idea and visualize it in my head. So within the early stage right here, I needed to see how the software works with concepting concepts, and distinction that with my regular course of. I needed to see how the picture creator would visualize what I’m pondering.

You’ll be able to see that the software is utilizing plenty of photo-reference pictures, however typically my work shouldn’t be as stiff—it’s extra summary. So I began making an attempt to push it to go to the specifics that I used to be on the lookout for visually.

Jon Han: Right here I shifted away from bridges and tried to push for a extra summary picture and a digital impact. I preferred the shapes of the panels and the digital really feel. But it surely regarded an excessive amount of like a sci-fi film. I wanted a happier temper.

Jon Han: This time, I added in ‘lush panorama,’ which made the picture extra atmospheric, not so darkish and digital. However the temperature of the image was a little bit too cool. I put within the phrase “heat,” and it got here out like every little thing was on fireplace. Very like every other software, like Photoshop or oil portray software program, there’s methods to some of these items, the place sure prompts create sure sorts of pictures. So there’s a studying curve to determining methods to do these minor changes to get the specifics you need.

Jon Han: This time I added my identify to the immediate to see if it might have the ability to reference my explicit type primarily based on my printed illustrations. Including my identify made the individual smaller and bought the fitting viewpoint, seeing the individual from the again. I do lots of compositions the place I do a small determine in a panorama. But it surely didn’t really feel hyper-referenced to my work, in order that was good.

There are lots of combined emotions amongst artists on AI-generated pictures, considerations whether or not these methods may very well be referencing artists’ work with out copyright. I attempted to steer away from something that appears like an artist I do know. However one factor about this software, in comparison with different AI instruments I’ve tried, is that it didn’t look like it was referencing somebody’s work. This was a superb endpoint for the concepting stage.

Jon Han: After these rounds of experimentation I had a transparent picture in my head of what I needed, so I created this idea sketch by hand. The entire image I needed to create was too complicated to speak to the picture creator. So relatively than deliver AI-generated pictures straight into this sketch, I made a decision to make use of the software [or the AI] in later levels to generate choices for explicit parts of the illustration, like timber or sky.

Usually, you would possibly do an web seek for images of issues like that, for reference. An artwork instructor as soon as taught me: by no means use the primary few pictures from an web search as your reference, as a result of everybody’s going to make use of these. So with the picture creator, it’s good that these are going to be generated a distinct method.

Jon Han: Earlier than creating the full-color illustration, I used the picture creator to check completely different sky colours. I additionally tried pixelated timber and wireframe timber, on the lookout for one thing fascinating that doesn’t have an excessive amount of of an AI really feel to it. This was all experimentation, discovering limitations, and experiments with management.

---

Han says that for lots of people, AI picture creators are “sort of like a shock field.” In his expertise, they’re most helpful once you don’t know precisely what you need. In a case the place a creator already has the picture they need of their head, a picture generator will be good for iterating particular parts.

We’re solely simply starting to see how next-generation AI can match into the artistic course of, and it’ll definitely look completely different for each artist and creator. From the prompts folks use to information the instruments to the diploma that AI does—or doesn’t—affect the ultimate product, every little thing is up for exploration.