Visions of the future? (aka playing around with Adobe's Generative Fill AI)

I spent a couple of hours over the weekend playing with Adobe's new Generative Fill tool. I'm impressed.

Although I'm not a big fan of Adobe, I recently broke down and bought a subscription so I could continue using Lightroom to store, manage, and edit my digital photographs (I tried open-source alternatives and they just don't compare). The subscription comes with a license to use Photoshop, which I've never used before...

...until this past weekend.

Last week, I started seeing videos on Instagram about the crazy new AI feature added to Photoshop. Seeing examples of what the Generative Fill tool could do, I was immediately curious to explore whether this could be applied in my PhD research, which will be looking at the images we use to visualise future food systems within the context of climate change.

While it's true we can't take pictures of the future, perhaps AI can help us create "images" or "visions" of imagined futures?

After watching a handful of YouTube videos on how to use the tool, I downloaded the beta version of Photoshop and started playing around. While AI image generators like Midjourney or DALL·E create entirely new images based on text prompts, the new Photoshop tool allows you to use text prompts to manipulate pre-existing images to create something new.

For my starting image, I chose this image of the rolling hills of Dartmoor that I shot before sunrise on a wild camping trip in the Spring of 2021. Using Photoshop's Generative Fill tool, I was able to turn this picture...

...into this.

Pretty impressive, right? I was shocked at the speed and quality of the AI generator, which takes just a few seconds to come up with three options based on the prompts you provide. Not only can you "expand" the image (e.g. I increased the size of the foreground), but you can (more or less) seamlessly add elements into the photo that were not there to begin with.

For this image, I used prompts such as "forest covered hills," "wind turbines on the horizon," and "two highland cows grazing in the distance," amongst others. Many of the added elements took a bit of trial and error, with multiple rounds of generation and slightly altered prompts before finding an option that fit the scene.

There were things the AI definitely struggled with – every time I tried adding "trees" or a "forest" into the valley, all it would give me were buildings or other human-made structures. It also struggled with people and animals, which may limit its applicability in certain research contexts.

As it was my first time using the tool, the original "future vision" took me a while to create, but these subsequent images, however, each required under 10 minutes of playing around with the Generative Fill tool. For anyone who is interested in trying this, the tool is very easy to use even with no prior Photoshop experience.

These images are far from perfect, but I hope they can show how this tool can use the same starting image to visualise a range of imagined "futures," both dystopian and utopian.

While the images I created are based simply on my imagination, the tool could also be used to create evidence-based "visions" of the future for use in research projects. There are a number of ways this tool could (perhaps) be used in research, which I'd love to hear your thoughts on:

• Researcher driven photo-elicitation could be used to explore how participants react to different "scenarios" created using AI-edited images of local areas/meaningful places.

• After a basic tutorial, participants could be encouraged to play directly with the AI tool in Photoshop, creating an image based on a certain scenario described by the researcher. The images created by participants could act as a "dataset" for visual analysis.

• Images of future scenarios could be co-created with key stakeholders or communities, through iterative and creative processes.

Here are a few other "before and after" images I created using photographs I've taken over the last couple of years (the photo on the left is the unedited original, while the one on the right is edited in Photoshop).

Using Adobe's Generative Fill tool, the images on the right are composed of both "real" and AI elements. They are, therefore, not entirely AI-generated, but are equally not "real" any more. While image editing is, of course, not new, this tool makes it even easier to manipulate images, raising the question (for me, at least): how will we ever know if an image is "real" or not?

While Adobe claims it is not using stolen images to teach its AI, there are still a number of ethical discussions to be had around using this kind of tool (in research and more broadly). For example, how will biases in the AI be reflected in the images we create?

I'd love to hear your thoughts on this! Have you played with the new Generative Fill tool in Photoshop? What are your first impressions of AI tools like this? Do you think they could make for valuable tools in researching climate visuals?