Obliviate: Erasing Concepts from Autoregressive Image Generation Models
Obliviate effectively erases concepts in autoregressive image generation, reducing nudity detection on RAB benchmark from 91.58% to 3.15%.
Key Findings
Methodology
Obliviate employs KL-based supervision over visual token distributions, trajectory-level updates, and aligned visual prefixes for concept erasure in autoregressive image generation. It introduces a pseudo-unconditional branch with a smooth KL objective to stabilize target construction.
Key Results
- On the RAB benchmark, Obliviate reduced nudity detection from 91.58% to 3.15% while preserving overall model utility.
- Obliviate outperformed existing methods in removing explicit content, graphic violence, and branded imagery across Liquid, Emu3-Gen, and Janus-Pro models.
- Full trajectory updates enabled rapid and effective concept erasure within 20 steps.
Significance
This research fills the gap in concept erasure for autoregressive image generation, providing an effective method to remove unsafe or legally sensitive concepts. Obliviate offers a flexible and persistent way to edit model behavior while maintaining generation quality.
Technical Contribution
Obliviate resolves the dynamic balance between autoregressive sequence modeling and visual generation through aligned visual prefixes and full trajectory KL supervision, achieving a technical breakthrough in concept erasure for autoregressive image generation.
Novelty
Obliviate is the first method to successfully achieve concept erasure in autoregressive image generation, introducing a pseudo-unconditional branch and full trajectory updates to address instability in previous methods.
Limitations
- Obliviate may experience performance degradation when handling very complex visual concepts, which may require more sophisticated model adjustments.
- Parameter tuning across different models may require additional effort.
Future Work
Future research could explore the application of Obliviate in larger-scale multimodal models and investigate its effectiveness in real-time generation systems.
AI Executive Summary
The widespread adoption of generative AI models has raised concerns about misuse, particularly in generating unsafe or disturbing imagery. Existing concept erasure methods have primarily focused on diffusion models, leaving autoregressive image generation largely unexplored. Obliviate introduces KL-based supervision over visual token distributions, trajectory-level updates, and aligned visual prefixes to achieve concept erasure in autoregressive image generation. Experimental results demonstrate that Obliviate outperforms existing methods in removing explicit content, graphic violence, and branded imagery while preserving overall model utility. This research fills the gap in concept erasure for autoregressive image generation, providing a new solution for enhancing the safety of generative models.
Deep Analysis
Background
In recent years, image generation has advanced rapidly, achieving remarkable realism and controllability. However, as image generators become more capable, the risk of generating unsafe or legally sensitive concepts increases. Concept erasure aims to remove specific unsafe capabilities from pre-trained generative models while preserving their generation quality.
Core Problem
While concept erasure research has matured in diffusion models, corresponding methods for autoregressive image generation remain underexplored. With the growing importance of autoregressive models in vision-language unified architectures, there is an urgent need for effective methods to address this safety issue.
Innovation
Obliviate introduces a pseudo-unconditional branch and full trajectory KL supervision to resolve the dynamic balance between autoregressive sequence modeling and visual generation, achieving a technical breakthrough in concept erasure for autoregressive image generation.
Methodology
- �� Use KL-based supervision over visual token distributions to ensure generated image tokens steer away from unsafe concepts.
- �� Implement trajectory-level updates to leverage supervision signals over the entire generation path.
- �� Align visual prefixes to stabilize target construction and identify tokens responsible for sustaining undesired concepts.
Experiments
Evaluated on three autoregressive text-to-image models: Liquid, Emu3-Gen, and Janus-Pro, covering the erasure of explicit content, graphic violence, and branded imagery. Benchmarks like RAB and Unbranding were used to ensure effectiveness and robustness.
Results
On the RAB benchmark, Obliviate reduced nudity detection from 91.58% to 3.15%. It outperformed existing methods in removing explicit content, graphic violence, and branded imagery across all tested models.
Applications
Obliviate can be applied in any scenario requiring safe image generation, including social media content moderation, advertising generation, and automated design.
Limitations & Outlook
While Obliviate excels in concept erasure, it may experience performance degradation when handling very complex visual concepts. Additionally, parameter tuning across different models may require extra effort.
Plain Language Accessible to non-experts
Imagine a factory that produces a variety of products. Sometimes, these products might include some inappropriate ones, like images with unsuitable content. Obliviate acts like an intelligent quality inspector that identifies and removes these inappropriate products without affecting the production of other normal products. By thoroughly checking the products, Obliviate ensures that each one meets safety standards.
ELI14 Explained like you're 14
Hey there! Imagine you're playing a game with a character creator that can make all sorts of characters. Sometimes, you might not want certain characters to appear, like those that aren't quite appropriate. Obliviate is like a super-smart helper that can remove those unwanted characters while keeping the other fun ones. Isn't that cool?
Glossary
Autoregressive Model
A model that generates sequences by predicting the next token. Used in image generation to predict the next visual token.
Used to predict the next visual token in image generation.
KL Divergence
A measure of difference between two probability distributions. Used to supervise visual token distributions.
Used to supervise visual token distributions.
Concept Erasure
The process of removing specific unsafe concepts from generative models.
Ensures the safety of generated images.
Visual Prefix
The initial sequence of visual tokens used in image generation.
Used to align conditional and unconditional branches.
Full Trajectory Update
A method of updating supervision signals over the entire generation path.
Improves the efficiency and effectiveness of concept erasure.
Open Questions Unanswered questions from this research
- 1 How can Obliviate be applied in larger-scale multimodal models?
- 2 What is the effectiveness of Obliviate in real-time generation systems?
Applications
Immediate Applications
Social Media Content Moderation
Use Obliviate to ensure generated images do not contain inappropriate content, enhancing platform safety.
Long-term Vision
Automated Design
Apply Obliviate in advertising generation and design to ensure generated content meets brand and legal standards.
Abstract
The widespread adoption of generative AI models has intensified concerns about misuse, including the creation of unsafe or disturbing imagery. To mitigate such issues, several concept erasure approaches have been proposed to remove harmful content from multimodal generative models. Yet concept erasure for autoregressive image generation remains largely unexplored, despite the growing relevance of these models in recent trends toward unified multimodal architectures. In this work, we fill this gap by introducing Obliviate, a guidance-based concept erasure method for autoregressive image generation. Our method builds on three key design choices: KL-based supervision over visual token distributions, trajectory-level updates over full autoregressive rollouts, and aligned visual prefixes for stable target construction. We evaluate Obliviate on three state-of-the-art autoregressive text-to-image models, Liquid, Emu3-Gen, and Janus-Pro, covering the erasure of explicit content, graphic violence, and branded imagery. Obliviate consistently outperforms current alternatives, reducing nudity on the defensive RAB benchmark from 91.58 to 3.15 while preserving overall model utility.