Training and Critiquing AI through Art with Tiffany Calvert

Artist Tiffany Calvert is busy wrapping up a day in her studio, putting the finishing touches on a largescale still life oil painting. In 2025, it might seem like an anachronism, something from another century. The catch: the image was created in part by a custom AI diffusion model.

Calvert, who chairs the Sam Fox School’s MFA in Visual Art program, has always been interested in blending old and new techniques. She believes there’s value in exploring historical traditions in modern contexts, asking what paintings and other artworks mean in a world of digital images and what they can tell us about perception and image commodity.

“I chose abstract painting as a mode on purpose — its intention is to interrupt an image and make it harder to look at, so that as a viewer, you have to think about your own process of looking,” she says. “As we’re flooded with images all day, if an abstract painting can ask you to slow down for a minute and feel yourself figuring out what you’re seeing, that’s so important to understanding and perceiving the truth, believability, and reliability of images.”

Calvert started focusing on Dutch and Flemish still life paintings in 2014. “They’re so incredibly detailed, that in some ways they become abstract — they’re no longer realistic,” she says. She gathered some 650 images of still life paintings from various internet archives and entered them into a StyleGAN, or style-based generative adversarial network. From there, she generated still life images to use as a springboard in her own work that involves painting abstractly in response to and over top of the still life.

“I’ve been using digital media in parallel to my artistic practice since my undergraduate degree in the ’90s,” she says. “At the time, I was developing my own artistic abilities at the same time as early versions of the internet and Photoshop 1 were coming out. Because I acquired that knowledge on my own, I felt like I had a solid understanding of the behind-the-scenes operations and how the apps actually worked. That continues to be very important to me.”

Then, a roadblock. Just as Calvert was arriving at WashU in fall 2024, the program she had been using changed its protocol and no longer supported her need for precise control over the image set. She explored other programs, but recognized she wasn’t going to progress without technical help due to the limitations of publicly available AI applications and her own coding knowledge.

“I was so happy to be at WashU at that moment, because there is real intent and motivation behind putting together teams for interdisciplinary research,” she says. Chad Henry, the school’s director of research and innovation, connected her to Associate Professor Ulugbek Kamilov and doctoral candidate Shirin Shoushtari in the McKelvey School of Engineering. “Bek and Shirin were already working on a custom diffusion model, and it was a very fortunate overlap in my methodologies and their project,” Calvert says.

The model the McKelvey team is developing can work with smaller data sets — like Calvert’s set of 650 images — with higher accuracy. The model doesn’t collage bits and pieces from different images, but rather invents a truly new image.

“Our work with Tiffany has been very valuable because we get real world, expert feedback on how our model is performing,” Shoushtari says. She went on to note that the collaboration with Calvert revealed to them that, in addition to viewing each generated image as a whole, they needed to focus on very small elements of each image to further the model’s ability. “While this exact model will only be applicable to Tiffany’s data set,” Shoushtari says, “the tools are transferable, so we may be able to use what we’ve learned here in other applications such as medical imaging.”

The team also hopes that the collaboration leads to a platform for artists to use to generate images with their own data sets. “So many AI platforms are taking control and artistic direction out of the hands of the artist,” Calvert says. “We hope to create something that gives back that control and precision.”

“A diffusion model develops an image by inserting noise periodically, systematically, and then the model learns to generate an image out of that noise,” Calvert says. “Abstract painting does the same thing. It’s inserting interruptions, obfuscations, complications, something irresolvable into an image — because then it captures your ability to process what you’re seeing.” By viewing an image for a long time, Calvert believes that you can begin to understand your own perception.

For her own practice, Calvert was most interested in the shortcomings of the AI model that reveal the very human biases in datasets, and therefore the dangers that AI will replicate and magnify those biases. “In order to visually critique the model, it’s more interesting to find those divisions, gaps, and failures in its processing,” she says. “Some of these AI-generated images we see online are so slick, so smooth… I think that abstract painting is uniquely positioned to criticize them, to push into the seams and deconstruct them to look at images more carefully,” she says.

Once the AI model has created a new image for Calvert to paint over, she gets to work preparing. A key part of her process is applying a custom-designed vinyl stencil, covering parts of the image while leaving openings for her to paint over. “I create the stencil to be reminiscent of cubism and desktop computer windows layered over each other,” she says. Once the canvas print of the image arrives, she stretches it on a frame, applies the vinyl, and spends a few hours mixing paint. Then, she has a single day to paint abstractly in response to the image and remove the vinyl before the paint dries.

“I find the boundary of a single day useful, because it means I can’t lose confidence in myself or the painting,” Calvert says. “I have to respond in a single, very concentrated session like this, and either it works or it doesn’t.”

As she begins to peel the vinyl, Calvert explains the value of prolonged looking. “If you stay with an image for a long time, you begin to look at it more critically, which has a big impact on how we perceive the world around us.”