Research Talk - Generative Modeling Via Drifting
Diffusion and flow-based models won the last era of generative AI, but they pay for their quality with dozens, sometimes hundreds of inference steps. Mingyang Deng and his collaborators at MIT and Harvard propose a different approach: Drifting Models, which evolve the generated distribution during training so that inference collapses to a single forward pass.
If you're working on research or building in world models, physical AI, simulation, or generative modeling, this paradigm directly attacks the inference bottleneck most of these systems share. Please come join us a research talk by Mingyang Deng from MIT on his paper: Generative Modeling via Drifting, followed by a mingling session afterward.
About the speaker
Mingyang Deng is pursuing PhD at MIT advised by Dr. Kaiming He. His work focuses on the foundations of generative modeling: building direct, end-to-end frameworks like Drifting Models and Mean Flows (NeurIPS 2025 Oral), and he co-authored Autoregressive Image Generation without Vector Quantization (NeurIPS 2024 Spotlight). Before his PhD he did his undergrad at MIT with a double major in Math and CS. He's also one of the most decorated competition students of his generation: 1st place at the ICPC World Finals, Putnam Fellow, gold and 1st place at IOI, and gold at IMO.
We'll gather 40-50 researchers and founders to go through the key sections of the research paper, followed by a networking session afterwards.
Hosted by 1943
1943 is a community for the researchers and builders who value intellectual daring over networking events. We host small group, closed-door roundtables and dinners across the Bay Area for those pushing the frontier. Follow us on Linkedin or X for future research talks and roundtable discussions!
Sponsored by Elorian
Elorian is building the foundation of visual reasoning. We believe building systems that natively understand and reason through the visual medium the way humans do is the most important work in AI today. It is also a critical step toward the future. If you do too, come and talk to us.