This work proposes to erase concepts by directly manipulating the model's internal representations by operating exclusively on the sparse text-branch tokens and leveraging the straight sampling trajectory of rectified flow, achieving effective concept erasure with negligible overhead and without any training.
Abstract
Multimodal Diffusion Transformers (MM-DiTs) have demonstrated remarkable text-to-image generation performance, surpassing traditional U-Net-based diffusion models. Nevertheless, their powerful generative capabilities also raise significant safety concerns, as they may generate sensitive or inappropriate content. While existing concept erasure methods aim to mitigate such risks, most require modifying model parameters, which are often architecture-specific and impractical for deployed larger models. Several tuning-free approaches face challenges when applied to advanced large-scale MM-DiTs due to their deeply embedded knowledge, broad semantic space, and context-dependent text encoders. To address these challenges, we propose to erase concepts by directly manipulating the model's internal representations. Our key insight, derived from an in-depth analysis of MM-DiT's block-wise generative roles, is that text-conditioned semantic representations are most salient in the middle blocks of MM-DiTs. Based on this, we extract representations of an unwanted concept and a desirable safe one from the middle block, construct a steering vector from their difference, and inject this single vector into consecutive early and middle blocks. By operating exclusively on the sparse text-branch tokens and leveraging the straight sampling trajectory of rectified flow, our method achieves effective concept erasure with negligible overhead and without any training. Extensive experiments across MM-DiT models demonstrate that our method achieves state-of-the-art performance in erasing diverse concepts, enables effective control over the final output, and remains robust to adversarial attacks.
Text-to-image diffusion models can be misused to generate harmful content through adversarial or paraphrased prompts that bypass built-in safety mechanisms. Existing concept erasure methods often suffer from limited robustness against adversarial prompts, degradation of benign generation quality, or reliance on inference-time interventions that introduce persistent computational overhead. To address these limitations, we formulate concept erasure as a domain alignment problem in the text representation space. We propose a lightweight Text Encoder Alignment framework (TEA) that fine-tunes only the text encoder while keeping the generative backbone fully frozen. Given concept--anchor prompt pairs, our method trains a discriminator to distinguish token-level representations of concept-containing prompts from those of safe anchor prompts, while updating the text encoder to make these representations indistinguishable. TEA introduces zero inference-time overhead and requires only a small number of fine-tuning steps, making it highly efficient to deploy at scale. Despite this efficiency, TEA achieves state-of-the-art erasure robustness against black-box and white-box adversarial attacks on Stable Diffusion v1.4, while preserving generation quality on benign prompts. Furthermore, TEA is model-agnostic and achieves the lowest attack success rate on Stable Diffusion v3.5, extending concept erasure to a Rectified Flow Transformer architecture with T5 conditioning where prior methods remain largely unexplored. Code is available at \href{https://github.com/alirezafarashah/TEA.git}{https://github.com/alirezafarashah/TEA.git}
UDT is proposed, a U-Net diffusion transformer that combines the representation power of DiTs with the encoding-decoding benefits of U-Nets, through data-adaptive token merging for downsampling and upsampling, while preserving the DiT token dimension.
Junno Yun, Y. Alçalar, Mehmet Akçakaya· 0 citations
Text-to-image diffusion models have achieved remarkable progress in image synthesis, but their potential misuse for generating unauthorized or harmful content has raised growing safety concerns. This has created an urgent need for safe diffusion-based image generation methods that can selectively suppress sensitive concepts while preserving the model’s general generative capability. Existing concept erasure approaches typically rely on either model fine-tuning or closed-form editing. However, they often suffer from two major limitations: (1 insufficient or excessive erasure, where the former fails to suppress target concepts and the latter disrupts benign semantics; and (2 degradation of non-target concepts, where removing target concepts undermines the generation of unrelated concepts, especially in multi-concept scenarios. To address these issues, we propose the Singular Value Eraser (SVEraser), a lightweight concept erasure module that removes specific concepts by optimizing singular-value offsets of weight matrices. Operating in a compact yet expressive singular-value space, SVEraser enables precise concept removal while reducing side effects on unrelated content. Moreover, once trained for different concepts, multiple SVErasers can be flexibly combined for multi-concept erasure. To further reduce interference, we introduce an eraser activation mechanism that adaptively selects the appropriate SVErasers during inference based on the input prompt. Extensive experiments on copyrighted objects, artistic styles, and explicit content demonstrate that our method achieves accurate target concept removal while preserving non-target semantics, providing a practical and reliable solution for safe diffusion-based image generation.
Xiaoyu Geng, Shuaixiong Hui, Yuxin Wang et al.· IEEE Transactions on Image P...· 0 citations
State-of-the-art flow based text-to-image (T2I) models exhibit remarkable generative abilities but remain vulnerable to producing unsafe content. Prior safety efforts range from concept erasure and prompt filtering to classifier-based gating. However, simple techniques like parameter efficient adaptations of the models easily bypass such guardrails. We introduce a unique principled approach that achieves safety by regulating the model's attention dynamics through inference-time introspection, exhibiting intrinsic robustness. Our method analyzes and rebalances attention activations throughout image synthesis, steering generations away from unsafe concepts while preserving semantic alignment. This introspective control ensures safety of deployed models. Across standard and adversarial safety benchmarks, our approach achieves remarkable safety scores while maintaining or even improving alignment and perceptual quality. Our results reveal that attention-space regulation offers a considerably more promising path to safer diffusion transformer based image generation than the existing concept erasing mechanism.Our code can be accessed at https://basim-azam.github.io/iam/
BiVidGen, a hybrid framework where an MLLM first generates semantic visual tokens and a DiT renders videos conditioned on both text and these tokens via multi-layer cross-attention is proposed.
Yanbo Ding, Yijia Fan, Caihua Shan et al.· 0 citations
Though zero-shot adapters excel in image personalization, they often encounter significant challenges in multi-reference personalized generation, specifically failing to precisely adhere to the spatial layouts described in text prompts and suffering from feature leakage between reference images. To address these two challenges, we propose RIG (Regional Image-prompt Generation), a novel training-free framework. For the first challenge, leveraging Multimodal Large Language Models (MLLMs), we introduce a layout planning binder. Leveraging Chain-of-Thought (CoT) reasoning, this module infers and generates precise global layouts from text prompts, while simultaneously binding reference images to their corresponding regions. For the second, we introduce a satially decoupled diffusion mechanism that isolates feature streams during attention computation. By injecting reference features exclusively into designated regions, this mechanism effectively prevents feature interference between reference images. Extensive experiments demonstrate that RIG significantly outperforms state-of-the-art adapter methods in terms of both personalization fidelity and text-layout alignment.
Junhao Feng· Poster Volume 0008 The 2026...· 0 citations