This repository is a curated collection of research papers focused on the alignment of diffusion models, primarily for text-to-image generation. It serves as a valuable resource for researchers and practitioners interested in making diffusion models adhere to human preferences and instructions. The collection aims to be comprehensive, covering various alignment techniques, benchmarks, and fundamental concepts.

How It Works

The repository organizes papers into categories such as "Alignment Techniques," "Benchmarks and Evaluation," and "Fundamentals." It highlights key methods like Reinforcement Learning from Human Feedback (RLHF), Direct Preference Optimization (DPO), and various prompt optimization strategies. The collection also includes foundational statistical and machine learning concepts relevant to preference modeling.

Quick Start & Requirements

This is a curated list of papers and does not involve code execution. The primary requirement is access to research papers, many of which are linked via PDF.

Highlighted Details

• Comprehensive coverage of alignment techniques including RLHF, DPO, and prompt optimization.
• Extensive lists of benchmarks and evaluation datasets for assessing model alignment.
• Inclusion of fundamental theoretical papers on preference modeling and reinforcement learning.
• Links to a survey paper, "Alignment of Diffusion Models: Fundamentals, Challenges, and Future," for a deeper dive.

Maintenance & Community

The project is maintained by the xie-lab-ml community, with contributions welcomed for corrections and suggestions. The survey paper lists authors from various institutions, indicating a collaborative research effort.

Licensing & Compatibility

The repository itself is a collection of links to research papers. The licensing of the individual papers is determined by their respective publishers or preprint servers. Compatibility for commercial use or closed-source linking depends on the licenses of the cited works.

Limitations & Caveats

This repository is a literature survey and does not provide executable code or pre-trained models. The rapid pace of research means the collection may not be exhaustive or entirely up-to-date with the very latest publications.