GeoClean: Training-Free Concept Erasure in Rectified Flow via Posterior-Competition Stabilization
Abstract
Flow Matching (FM) models enable efficient few-step image generation, making it important to suppress unsafe or copyrighted concepts without retraining. Yet guidance developed for diffusion-aligned settings can degrade image quality and destabilize editing when transferred to practical rectified-flow models. We explain this failure through posterior competition and late-time amplification. In competition-prone regions, branch reweighting governs the concept-conditioned posterior response, so local guidance can incur correction mismatch. The rectified affine path then amplifies this mismatch near terminal time through mean-to-velocity conversion and increasing local sensitivity. Motivated by this analysis, we propose GeoClean, a training-free framework combining Competition-Aware Lookahead Evaluation (CLE) and Amplification-Controlled Correction Smoothing (ACS). CLE stabilizes where the correction is evaluated, while ACS regulates its temporal evolution before injection. Experiments show a 45.9% average ASR reduction while preserving image fidelity, semantic alignment, and few-step efficiency.