Obliviate: Erasing Concepts from Autoregressive Image Generation Models

The widespread adoption of generative AI models has inten-sified concerns about misuse, including the creation of unsafe or disturb-ing imagery. To mitigate such issues, several concept erasure approacheshave been proposed to remove harmful content from multimodal gen-erative models. Yet concept erasure for autoregressive image generationremains largely unexplored, despite the growing relevance of these modelsin recent trends toward unified multimodal architectures. In this work,we fill this gap by introducing Obliviate, a guidance-based concepterasure method for autoregressive image generation. Our method buildson three key design choices: KL-based supervision over visual tokendistributions, trajectory-level updates over full autoregressive rollouts,and aligned visual prefixes for stable target construction. We evaluateObliviate on three state-of-the-art autoregressive text-to-image mod-els, Liquid, Emu3-Gen, and Janus-Pro, covering the erasure of explicitcontent, graphic violence, and branded imagery. Obliviate consistentlyoutperforms current alternatives, reducing nudity on the defensive RABbenchmark from 91.58 to 3.15 while preserving overall model utility.