What's next in generative modeling?
The field of generative modeling has seen remarkable progress, with innovative methods redefining how we model complex data distributions. However, many challenges remain: Among others, the computational burden and sample complexity of high-quality methods, model evaluation and interpretability, and the flexibility to apply and adapt models to arbitrary data modalities with ease.
The Novel Approaches to Generative Modeling workshop aims to provide a platform for researchers and practitioners to showcase approaches that push the boundaries of generative modeling. We invite contributions that address current challenges through fresh perspectives, whether by introducing novel architectures, leveraging insights from adjacent disciplines, or rethinking the foundational assumptions of generative modeling.
The workshop welcomes submissions across a range of themes designed to capture emerging trends and persistent challenges in generative modeling. These tracks serve as organizing principles, not strict categories, and we encourage creative, boundary-pushing contributions.
Novel architectures This track highlights recent advances in model design that expand the expressive capacity, (data-)efficiency, or adaptability of generative models. We’re particularly interested in approaches that depart from established paradigms or reimagine the structure of generative systems in meaningful ways.
Novel training paradigms Training remains one of the most delicate aspects of generative modeling. This track focuses on methods that propose alternative learning strategies—whether by stabilizing objectives, improving sample efficiency, or addressing challenges like convergence and mode collapse.
Evaluation-driven innovations Progress in generative modeling is tightly coupled with how we define and measure success. This track invites work where insights from evaluation—whether metric-based, human-centered, or task-specific—lead to deeper understanding or new modeling techniques.
Application-driven innovations Generative models are increasingly deployed in settings with complex constraints and requirements. This track focuses on work where the demands of a specific application inspire new methods, formulations, or modeling considerations.
Why does this not work? Failure is often more instructive than success. This track provides space for thoughtful analysis of approaches that fall short—whether due to limitations in theory, implementation, or assumptions.
Felix Draxler, Postdoc @ UC Irvine
Sander Hummerich, PhD Candidate @ Heidelberg University
Stephan Mandt, Associate Professor of Computer Science and Statistics @ UC Irvine
Stefan Radev, Assistent Professor @ Rensselaer Polytechnic Institute
Armand Rousselot, PhD Candidate @ Heidelberg University
Farrin Marouf Sofian, PhD Candidate @ UC Irvine
Julia Vogt, Assistant professor in Computer Science @ ETH Zurich