This repository is a curated list of resources for discrete diffusion models, targeting researchers and practitioners in machine learning, natural language processing, and computational biology. It provides a structured overview of papers, code, and introductory materials, facilitating the exploration and advancement of discrete diffusion techniques for generative tasks.

How It Works

The list categorizes papers based on their core technical contributions, such as the type of noise used (discrete vs. Gaussian), specific methodologies like discrete flows and samplers, guidance mechanisms, custom noise processes, theoretical underpinnings, and application areas. This organization allows users to quickly identify relevant research within the rapidly evolving field of discrete diffusion models.

Quick Start & Requirements

This is a curated list, not a runnable codebase. Resources linked within the list may have their own installation and execution requirements.

Highlighted Details

• Comprehensive coverage of discrete diffusion models, including papers from major conferences like NeurIPS, ICML, ICLR, and ACL.
• Categorization spans various aspects: noise types, samplers, guidance, custom noise processes, theory, and applications.
• Includes links to abstracts and available code for most listed papers, enabling direct access to implementations and further details.
• Features introductory materials and surveys for those new to the field.

Maintenance & Community

Maintained by Subham Sahoo, Yingheng Wang, and Yair Schiff. Contributions are welcomed via pull requests, with a preference for chronological order and precedence for accepted papers.

Licensing & Compatibility

The licensing of individual papers and code repositories linked within this list varies. Users should consult the specific licenses of each resource.

Limitations & Caveats

This is a list of resources, not a unified framework or library. Users must individually assess the quality, applicability, and licensing of each linked paper and its associated code. The field is rapidly evolving, and the list may not be exhaustive.