Reproduce “An LLM Erases a Concept by Editing Diffusion Weights Directly, Without Gradients” 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
Javad Taghia
The authors present an LLM agent that edits text-to-image diffusion-model weights without gradient-based training to reduce imitation of a named artist. The agent uses model inspections and bounded feedback to choose tensor edits. On held-out prompts, it lowers Van Gogh detections more than random search on two models, but CLIP-based tests cover one style and do not establish complete erasure.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.