ScaleErasure: Inference-Time Minimal Intervention for Precise Concept Erasure in Next-Scale Autoregressive Image Generation

2026-06-28

AI Summary

This work addresses the challenge of precisely erasing unsafe concepts in next-generation scale-wise autoregressive image generation, where early-scale semantic compression causes severe entanglement between harmful content and irrelevant features, making it difficult to remove sensitive information without compromising general generative capabilities. To this end, we propose ScaleErasure—the first inference-time concept erasure method tailored for this paradigm. By performing two lightweight forward passes targeting both unsafe concepts and their safe counterparts, ScaleErasure jointly selects and steers target logits across scale, token, and bit-channel dimensions, achieving precise removal with minimal intervention. Experiments demonstrate that ScaleErasure significantly outperforms existing baselines, effectively eliminating unsafe content while better preserving the model’s general-purpose generation ability.

Abstract

Concept erasure aims to prevent image generative models from producing unsafe content while preserving their general generative capability. Meanwhile, next-scale autoregressive (AR) image generation has recently emerged as a new generative paradigm characterized by next-scale prediction, for which concept erasure remains largely unexplored. In this paradigm, semantic information is highly compressed at early scales, leading to severe entanglement between unsafe and unrelated semantics. In this paper, we propose ScaleErasure, an inference-time concept erasure method that performs minimal intervention. ScaleErasure precisely selects and guides predicted logits that are most relevant to the unsafe concept, thereby enabling effective erasure under severe semantic entanglement. Specifically, ScaleErasure performs two additional forward passes conditioned on the unsafe concept and the corresponding safe concept, and leverages their outputs to guide the target logits away from unsafe concepts toward safe concepts. To enable precise and minimal intervention, logits selection and guidance are conducted across three dimensions: scales, tokens, and bit channels. Experiments demonstrate that ScaleErasure outperforms adapted baselines in the next-scale AR paradigm, achieving more precise concept erasure while largely preserving general generative capability. The code is available at https://github.com/coziiizz/ScaleErasure.

Problem

• concept erasure
• next-scale autoregressive image generation
• semantic entanglement
• unsafe content
• inference-time intervention

Innovation

• concept erasure
• next-scale autoregressive generation
• inference-time intervention
• semantic disentanglement
• logits guidance