GlamAI’s breakthrough research at ICML 2025 could push the industry forward in solving one of visual AI’s most frustrating problems
Anyone who’s experimented with AI image generators knows the frustration: you’re trying to create the perfect product shot or marketing visual, and suddenly the AI decides to give your model three arms or turn their face into something out of a funhouse mirror. It’s the kind of unpredictable behavior that makes businesses hesitant to fully embrace AI-generated content.
Now, researchers at California-based GlamAI think they’ve figured out why this happens—and more importantly, how to fix it.
At this year’s International Conference on Machine Learning (ICML), GlamAI’s Alexander Lobashev co-authored research that gets to the heart of AI’s visual “hallucination” problem. The paper, with the academic title “Hessian Geometry of Latent Space in Generative Models,” essentially maps out why AI sometimes goes completely off the rails when generating images.
“Think of it like this,” explains Paul Shaburov, GlamAI’s founder. “The AI has an internal compass it uses to navigate different visual possibilities. What we discovered is that this compass has dead zones—areas where a tiny adjustment in one direction can send you spinning wildly off course.”
The research reveals that AI’s visual “imagination”—what researchers call latent space—isn’t one smooth landscape but rather a series of distinct neighborhoods with treacherous borders between them. Cross one of these invisible boundaries, and your sleek product photo might suddenly sprout distorted features or impossible proportions.
“The erratic results aren’t random glitches,” Shaburov notes. “They’re actually fundamental features of how these models work. Once you understand that, you can start designing around it.”
For businesses, this matters enormously. A fashion brand can’t afford to have their latest campaign feature models with warped faces. An e-commerce company needs product images that look consistently realistic. When AI fails in these high-stakes scenarios, it doesn’t just look bad—it can seriously damage brand credibility.
GlamAI has been working on this problem from both ends. The company, which boasts over 1.3 million monthly active users and ranks among the top five photo editing apps on the App Store, combines cutting-edge research with practical solutions. Their platform blends open-source AI models with proprietary datasets and internal tools designed to keep generated images looking natural and consistent.
“Our clients range from individual Instagram influencers to major fashion houses,” Shaburov says. “They all need the same thing: visuals that look professional and authentic, not like they came from some AI fever dream.”
The research breakthrough comes at a crucial time for the visual AI industry. As more businesses integrate AI-generated content into their marketing strategies, the tolerance for unpredictable results is shrinking. Consumers are becoming savvier at spotting AI-generated content, and brands are increasingly concerned about the reputational risks of obvious AI artifacts.
What makes GlamAI’s approach particularly interesting is how they’re bridging the gap between academic research and real-world application. While many AI companies focus on either pure research or commercial products, GlamAI is betting that solving the fundamental problems will give them a lasting competitive advantage.
“This isn’t just about making our app work better,” Shaburov explains. “We’re trying to push the entire field forward. When visual AI becomes more reliable and predictable, everyone benefits—creators, brands, and consumers.”
The timing couldn’t be better. As AI-generated content becomes increasingly sophisticated, the line between human and machine creation continues to blur. But for businesses to fully embrace this technology, they need to trust that their AI tools won’t suddenly produce something embarrassing or off-brand.
If GlamAI’s research proves successful in practical applications, it could mark a turning point for the visual AI industry—the moment when AI-generated content becomes truly ready for prime time.
Copyright © 2025 California Business Journal. All Rights Reserved.
For California Business Journal Disclaimers, go to https://calbizjournal.com/terms-conditions/.
