Hugging Face Diffusers Ethical Guidelines

Hugging Face has published an ethical framework for the Diffusers library to ensure that technical decisions and community contributions are guided by a set of core values. This initiative aims to provide transparency regarding how the library is managed and to mitigate the potential negative societal impacts of diffusion model applications.

Core Ethical Guidelines

Hugging Face defines ethics as a continuous process of adaptation based on guiding values and concrete actions. The Diffusers library is governed by the following six principles:

  • Transparency: A commitment to explaining technical decisions and the management of pull requests to users.
  • Consistency: Ensuring stable project management and a consistent level of attention for all users.
  • Simplicity: Maintaining lean and coherent project goals to ensure the library remains easy to use and exploit.
  • Accessibility: Lowering the entry barrier for contributors, including those without deep technical expertise, to make research artifacts more accessible.
  • Reproducibility: Maintaining transparency regarding the reproducibility of datasets, models, and upstream code provided through the library.
  • Responsibility: A collective commitment among the community and team to anticipate and mitigate the risks and dangers associated with diffusion technology.

Safety Features and Mechanisms

Hugging Face and the community have implemented several mechanisms to promote the responsible deployment and evaluation of diffusion models:

Community and Content Governance

  • Community Tab: A dedicated space for collaboration and discussion on projects.
  • Tagging: Repository authors can use a "Not For All Eyes" tag to mark sensitive content.
  • Staged Releases: The ability to restrict access to specific repositories in sensitive situations, allowing authors to maintain more control over usage.
  • Licensing: The use of OpenRAILs licenses to balance free access with restrictions that ensure responsible use.

Technical Safety and Bias Evaluation

  • Bias Exploration: Hugging Face provides a dedicated Space to interactively demonstrate biases present in models like DALL-E and Stable Diffusion, encouraging further bias evaluation.
  • Safe Stable Diffusion: To address inappropriate degeneration in models trained on unfiltered web-crawled datasets, the library includes StableDiffusionSafe (based on the research paper Safe Latent Diffusion: Mitigating Inappropriate Degeneration in Diffusion Models).

Sources