TLDR: GlamAI team member Alexander Lobashev co-authored a paper presented at ICML 2025, titled ‘The Hessian Geometry of Latent Space in Generative Models.’ This research explains the unpredictable changes in AI image editing, contributing to a theoretical understanding of generative models’ unstable behavior and aiming for more reliable and realistic AI-generated images.
San Francisco, CA – GlamAI, a prominent platform in AI-powered visual marketing for the influencer economy and fashion industries, announced that its team member, Alexander Lobashev, co-authored a significant research paper presented at the International Conference on Machine Learning (ICML) 2025. The paper, titled ‘The Hessian Geometry of Latent Space in Generative Models,’ addresses a critical challenge in AI image editing: the phenomenon where seemingly minor adjustments can lead to sudden and drastic alterations in an image. This research posits that such unpredictable ‘jumps’ are not random glitches but rather an inherent characteristic of how these generative models operate.
The study, co-authored by a consortium of distinguished AI scientists from leading industry labs and academic institutions, offers a substantial contribution to the theoretical understanding of the unstable and unpredictable behaviors observed in generative models. In practical terms, the findings shed light on why two nearly identical input images can yield vastly different, and at times unrealistic, outputs. This inconsistency often manifests as distorted shapes, loss of identity, or undesired ‘hallucinations’ when applying filters, styles, or transformations using contemporary AI models, leading to user frustration and reduced trust in AI tools.
Paul Shaburov, Founder of GlamAI, commented on the significance of the research, stating, ‘This problem has been a major pain point in the visual AI space, particularly when realism and brand integrity are critical.’ He further emphasized that ‘The research Alexander contributed to helps build a theoretical foundation to explain and fix this issue—and it’s a great example of how GlamAI combines applied engineering with world-class research to move the entire industry forward.’
In essence, the research provides a geometric framework for understanding why models can unexpectedly ‘break’ or produce anomalous results during image manipulation. This insight is deemed a crucial step toward developing models that are inherently more stable and capable of delivering predictable, high-quality outcomes. It empowers teams like GlamAI’s to refine and control model behavior, ensuring realistic and anticipated editing results. The study holds particular relevance for fields such as generative photography, where users expect AI-enhanced visuals to maintain a natural appearance and preserve identity.
Also Read:
- Generative AI Poised to Transform Fashion Design and Trend Forecasting, New Research Reveals
- Hyper3D Launches Rodin AI Suite, Revolutionizing 3D Model Generation from Text and Images
GlamAI’s commitment to advancing both its product capabilities and the underlying scientific principles is evident in Lobashev’s contribution. The company has seen considerable success, reaching 1.3 million monthly active users and ranking among the top five most popular apps in the ‘Photo and Video Editing’ category on the App Store as of April 2025. This research is part of GlamAI’s ongoing efforts to deliver the realistic-looking generated images that their customers demand.
