Reproduce “Editing Implicit Assumptions in Text-to-Image Diffusion Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Hadas Orgad, Bahjat Kawar, Yonatan Belinkov
Researchers from Technion, Israel introduce TIME, an efficient method for editing implicit assumptions and biases within pre-trained text-to-image diffusion models. By modifying less than 3% of model parameters in under a second, TIME enables targeted alterations to model behavior, such as mitigating gender stereotypes or updating factual associations, while preserving overall generative quality.
Key insights
- The cross-attention layers' W_K and W_V matrices are critical and precise targets for modifying text-based implicit knowledge in text-to-image models.
- Targeting a small, specific subset of parameters (2.2% of the diffusion model) allows for highly efficient, localized edits that preserve the model's overall generative capabilities.
- Leveraging a closed-form solution for optimization enables near-instantaneous, single-step modifications, significantly improving efficiency over traditional iterative training methods.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.