# [MASK] is All You Need

**Authors:** Vincent Tao Hu, Björn Ommer  
**arXiv:** [2412.06787](https://arxiv.org/abs/2412.06787)  
**Date:** 2024-12-09  
**Summary page:** https://www.summarizepaper.com/en/paper/2412.06787v2/

## Key points

- Two prominent paradigms in generative models: Masked Generative Models and Non-Autoregressive Models
- Proposal of a novel approach bridging these paradigms using discrete-state models in vision domain
- Methodology involves comprehensive analysis across both types of models and redefines traditional discriminative tasks as an unmasking process within a discrete-state model
- Introduction of framework called Discrete Interpolants leading to state-of-the-art performance on various benchmarks
- Leveraging [MASK] in discrete-state models to bridge gap between different generative models and integrate generative and discriminative tasks seamlessly
- Potential extension to other approaches by utilizing discrete stochastic interpolants mentioned in related works
- Acknowledgment of contributions from Timy Phan, Moyang Li, and Owen Vincent for proofreading assistance and technical support
- Support received from various entities including the German Federal Ministry for Economic Affairs and Climate Action, Bayer AG, and the German Research Foundation (DFG)
- Gratitude expressed to the Gauss Center for Supercomputing for providing computational resources through NIC on JUWELS at JSC and HPC resources from the Erlangen National High Performance Computing Center (NHR@FAU funded by DFG)

## Summary

In the realm of generative models, two prominent paradigms have emerged: Masked Generative Models and Non-Autoregressive Models. In this study, we propose a novel approach that bridges these two paradigms using discrete-state models in the domain of vision. Our methodology involves a comprehensive analysis across both types of models and redefines traditional discriminative tasks as an unmasking process within a discrete-state model. Through our exploration and experimentation, we introduce a framework called Discrete Interpolants and achieve state-of-the-art performance on various benchmarks. By leveraging [MASK] in discrete-state models, we effectively bridge the gap between different generative models while seamlessly integrating generative and discriminative tasks. Our method has the potential to extend to other approaches mentioned in related works by utilizing discrete stochastic interpolants. We would like to acknowledge the contributions of Timy Phan and Moyang Li for proofreading assistance and Owen Vincent for technical support. This project has received support from various entities including the German Federal Ministry for Economic Affairs and Climate Action, Bayer AG, and the German Research Foundation (DFG). We also express gratitude to the Gauss Center for Supercomputing for providing computational resources through NIC on JUWELS at JSC and HPC resources from the Erlangen National High Performance Computing Center (NHR@FAU funded by DFG).