Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LURE: Latent Space Unblocking for Multi-Concept Reawakening in Diffusion Models

Mengyu Sun, Ziyuan Yang, Andrew Beng Jin Teoh, Junxu Liu, Haibo Hu, Yi Zhang

TL;DR

This paper tackles the vulnerability of concept erasure in diffusion models by reframing generation as an implicit function of text, model parameters, and latent states. It introduces LURE, a latent-space unblocking framework combining semantic re-binding for latent-space reconstruction, gradient field orthogonalization to prevent cross-concept interference, and LSIS with posterior verification to stabilize sampling. The approach enables simultaneous, high-fidelity reawakening of multiple erased concepts across general objects, IP concepts, and unsafe content while preserving non-erased semantics. The results imply that safe deployment of diffusion models requires mechanisms that disrupt latent generative knowledge, not just surface-level output suppression, to ensure robust safety.

Abstract

Concept erasure aims to suppress sensitive content in diffusion models, but recent studies show that erased concepts can still be reawakened, revealing vulnerabilities in erasure methods. Existing reawakening methods mainly rely on prompt-level optimization to manipulate sampling trajectories, neglecting other generative factors, which limits a comprehensive understanding of the underlying dynamics. In this paper, we model the generation process as an implicit function to enable a comprehensive theoretical analysis of multiple factors, including text conditions, model parameters, and latent states. We theoretically show that perturbing each factor can reawaken erased concepts. Building on this insight, we propose a novel concept reawakening method: Latent space Unblocking for concept REawakening (LURE), which reawakens erased concepts by reconstructing the latent space and guiding the sampling trajectory. Specifically, our semantic re-binding mechanism reconstructs the latent space by aligning denoising predictions with target distributions to reestablish severed text-visual associations. However, in multi-concept scenarios, naive reconstruction can cause gradient conflicts and feature entanglement. To address this, we introduce Gradient Field Orthogonalization, which enforces feature orthogonality to prevent mutual interference. Additionally, our Latent Semantic Identification-Guided Sampling (LSIS) ensures stability of the reawakening process via posterior density verification. Extensive experiments demonstrate that LURE enables simultaneous, high-fidelity reawakening of multiple erased concepts across diverse erasure tasks and methods.

LURE: Latent Space Unblocking for Multi-Concept Reawakening in Diffusion Models

TL;DR

This paper tackles the vulnerability of concept erasure in diffusion models by reframing generation as an implicit function of text, model parameters, and latent states. It introduces LURE, a latent-space unblocking framework combining semantic re-binding for latent-space reconstruction, gradient field orthogonalization to prevent cross-concept interference, and LSIS with posterior verification to stabilize sampling. The approach enables simultaneous, high-fidelity reawakening of multiple erased concepts across general objects, IP concepts, and unsafe content while preserving non-erased semantics. The results imply that safe deployment of diffusion models requires mechanisms that disrupt latent generative knowledge, not just surface-level output suppression, to ensure robust safety.

Abstract

Concept erasure aims to suppress sensitive content in diffusion models, but recent studies show that erased concepts can still be reawakened, revealing vulnerabilities in erasure methods. Existing reawakening methods mainly rely on prompt-level optimization to manipulate sampling trajectories, neglecting other generative factors, which limits a comprehensive understanding of the underlying dynamics. In this paper, we model the generation process as an implicit function to enable a comprehensive theoretical analysis of multiple factors, including text conditions, model parameters, and latent states. We theoretically show that perturbing each factor can reawaken erased concepts. Building on this insight, we propose a novel concept reawakening method: Latent space Unblocking for concept REawakening (LURE), which reawakens erased concepts by reconstructing the latent space and guiding the sampling trajectory. Specifically, our semantic re-binding mechanism reconstructs the latent space by aligning denoising predictions with target distributions to reestablish severed text-visual associations. However, in multi-concept scenarios, naive reconstruction can cause gradient conflicts and feature entanglement. To address this, we introduce Gradient Field Orthogonalization, which enforces feature orthogonality to prevent mutual interference. Additionally, our Latent Semantic Identification-Guided Sampling (LSIS) ensures stability of the reawakening process via posterior density verification. Extensive experiments demonstrate that LURE enables simultaneous, high-fidelity reawakening of multiple erased concepts across diverse erasure tasks and methods.
Paper Structure (22 sections, 5 theorems, 10 equations, 4 figures, 5 tables)

This paper contains 22 sections, 5 theorems, 10 equations, 4 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Theoretical Analysis
  4. Preliminaries and Problem Formulation
  5. Concept Entanglement Analysis
  6. Concept Erasure
  7. Method
  8. Problem Formulation
  9. Overview
  10. Latent Space Reconstruction
  11. Semantic Re-binding Loss.
  12. Gradient Field Orthogonalization Loss.
  13. Overall Loss Function.
  14. Latent Semantic Identification-Guided Sampling Strategy
  15. Experiments
  16. ...and 7 more sections

Key Result

Theorem 1

We define the reconstruction loss for concept $c_i$ as: where $x_0$ is sampled from the data distribution $p(x|c_i)$ for $c_i$. Then, let $c_i, c_j \in \mathcal{C}$ be two distinct concepts with $i \neq j$. The gradients of the two concepts exhibit non-orthogonality: Different concepts may become entangled in the latent space in the absence of the lack of explicit constraints during training. Th

Figures (4)

  • Figure 1: Concept illustrations of concept erasure and reawakening paradigms. (a) Original prompt–concept alignment enables direct semantic generation. (b) Concept erasure suppresses target concepts by blocking the original sampling trajectory. (c) Prompt-based methods seek alternative trajectories by optimizing or reformulating the prompts to bypass the erasure barrier. (d) Our method reconstructs the latent space to unblock the direct semantic trajectory.
  • Figure 2: Illustration of the proposed LURE framework, which consists of two core components: the latent space reconstruction and latent semantic identification-guided sampling.
  • Figure 3: Visualize comparison shows multi-concept reawakening results (Green box) on objects across UCE, RECE, and SPEED erasure methods (Red box) compared to original SD (Blue box).
  • Figure 4: Qualitative comparison of concept reawakening by LURE on Celebrity Identities and Intellectual Property Characters.

Theorems & Definitions (6)

  • Definition 1: Text-Visual Alignment Function
  • Theorem 1: Concept Entanglement
  • Lemma 1: Parameter Sharing
  • Theorem 2: Gradient Conflict
  • Theorem 3: Erasure-Induced Shift
  • Theorem 4: Concept Recoverability via Latent Reconstruction