The field of diffusion models is moving towards improved interpretability and security. Researchers are developing new methods to analyze and understand the internal workings of these models, such as the use of mechanistic interpretability techniques and novel visualization approaches. These advancements have the potential to increase trust in diffusion models and enable more effective steering of the generative process. Additionally, the community is acknowledging the importance of security in text-to-image diffusion models, with a focus on detecting and mitigating backdoor poisoning attacks. Noteworthy papers in this area include:

• The introduction of Diffusion Steering Lens, a novel approach for interpreting vision transformers.
• The proposal of REDEditing, a relationship-driven precise backdoor poisoning method for text-to-image diffusion models.
• The development of Prompt-Agnostic Image-Free Auditing, a scalable and practical solution for pre-deployment concept auditing of diffusion models.
• The application of Sparse Autoencoders to uncover human-interpretable concepts in diffusion models and demonstrate their causal effect on the model output.