EraseFlow: Learning Concept Erasure Policies via GFlowNet-Driven Alignment

TL;DR

EraseFlow introduces a novel concept unlearning framework for diffusion models that explores denoising trajectories with GFlowNets, effectively erasing harmful concepts while maintaining image quality and generalizing to unseen concepts.

Contribution

It is the first to formulate concept erasure as trajectory exploration in denoising paths using GFlowNets, avoiding brittle adversarial losses and retraining cycles.

Findings

Outperforms existing baselines in concept erasure tasks.

Effectively generalizes to unseen concepts.

Balances erasure effectiveness with prior preservation.

Abstract

Erasing harmful or proprietary concepts from powerful text to image generators is an emerging safety requirement, yet current "concept erasure" techniques either collapse image quality, rely on brittle adversarial losses, or demand prohibitive retraining cycles. We trace these limitations to a myopic view of the denoising trajectories that govern diffusion based generation. We introduce EraseFlow, the first framework that casts concept unlearning as exploration in the space of denoising paths and optimizes it with GFlowNets equipped with the trajectory balance objective. By sampling entire trajectories rather than single end states, EraseFlow learns a stochastic policy that steers generation away from target concepts while preserving the model's prior. EraseFlow eliminates the need for carefully crafted reward models and by doing this, it generalizes effectively to unseen concepts and avoids hackable rewards while improving the performance. Extensive empirical results demonstrate that EraseFlow outperforms existing baselines and achieves an optimal trade off between performance and prior preservation.