Forgetting is Competition: Rethinking Unlearning as Representation Interference in Diffusion Models

In three linesSurgUn, an unlearning method for text-to-image diffusion models, treats forgetting as controlled competition rather than direct deletion. Using target-gradient ascent and descent over semantically diverse distractors, it reduces erase-retain imbalance and limits collateral damage. Tested on Stable Diffusion v1.5, SDXL, and SANA-1.5.

Summary generated by Claude — human-verified