Yair Schiff
@SchiffYair
Remasking samplers tackle a key limitation of SoTA masked diffusion models: their inability to fix mistakes
This unlocks improved and more guidable generation and the benefits of scaling test-time compute!
Check out @Guanghan__Wang's thread for details about our new work: ReMDM
This unlocks improved and more guidable generation and the benefits of scaling test-time compute!
Check out @Guanghan__Wang's thread for details about our new work: ReMDM
Guanghan Wang@Guanghan__Wang · Mar 10, 2025We are excited to present ReMasking Diffusion Models (ReMDM), a simple and general framework for improved sampling with masked discrete diffusion models that can benefit from scaled inference-time compute.
Without any further training, ReMDM is able to enhance the sample quality
Without any further training, ReMDM is able to enhance the sample quality
GIF 0 5