Multiple Classes Erasure Using Superclass in Text-to-image Diffusion Models

Rajkumar, Pranav singh chib, Pravendra Singh

Keywords: Diffusion Model, Generative models, Text-to-image

TL;DR: We propose Context Graph Erasure, a framework that leverages structured scene knowledge through context graphs to guide precise and controllable concept erasure.

Abstract: Recent advances in text-to-image diffusion models have enabled highly realistic image synthesis, but they also raise concerns regarding the generation of unwanted or unsafe content. Existing erasure methods often struggle to remove target classes reliably when prompts are detailed, paraphrased, or adversarial, leading to incomplete forgetting and compromised fidelity of non-target content. To address these limitations, we propose Context Graph Erasure (CGE), a framework that leverages structured scene knowledge through context graphs to guide precise and controllable concept erasure. CGE constructs enriched representations of the visual scene by encoding objects, attributes, and relations into a learnable graph-based embedding, which is integrated with the text conditioning. A dedicated erasure module utilizes this enriched representation to suppress target superclass, while a cross-attention mechanism preserves the integrity of unrelated regions. Furthermore, an adversarial concept graph strategy allows the system to manage prompts that are phrased differently, maintaining consistent results. Extensive experiments demonstrate that CGE achieves superior erasure accuracy and preserves unrelated content with high fidelity, outperforming prior methods and providing a reliable, generalizable solution for multiple classes erasure in text-to-image models.

Primary Area: generative models

Submission Number: 10841