December 16, 2025
Dr Ioannis Kakogeorgiou, Researcher at the Institute of Informatics & Telecommunications at NCSR Demokritos (IIT), participated in the Neural Information Processing Systems Conference 2025 in San Diego, 2-7 December 2025.
The paper “Boosting Generative Image Modeling via Joint Image-Feature Synthesis”, that Dr Kakogeorgiou has co-authored, was accepted as a Spotlight, representing 3.2% of total submissions.
NeurIPS is one of the world’s leading interdisciplinary conferences in artificial intelligence and machine learning. This year, the conference received 21,575 valid paper submissions, of which 5,290 papers were accepted, with only 688 selected as Spotlight presentations.
The work presented in the paper introduces a latent-semantic diffusion framework that jointly models low-level image latents from a VAE and high-level semantic representations from self-supervised encoders such as DINO within a single diffusion process. By generating coherent image–feature pairs directly from noise, the approach improves image quality and accelerates training convergence with only minimal changes to standard Diffusion Transformer architectures.
