<ul><li>Masked diffusion models (MDM) are generative models for discrete data that use a partial masking scheme to prevent redundant computation.</li><li>The proposed method, Partial masking scheme (Prime), allows tokens to have intermediate states between masked and unmasked, improving the model's efficiency.</li><li>This approach enables the model to make predictions based on partially observed token information and enhances the denoising process.</li><li>The method shows superior performance on generative modeling tasks, achieving lower perplexity on text data and competitive FID scores on image data.</li></ul>

Beyond Masked and Unmasked: Discrete Diffusion Models via Partial Masking

Discover more