Diffusion Model Unlearning

Concept unlearning has emerged as a promising direction for reducing the risks of harmful content generation in text-to-image diffusion models by selectively erasing undesirable concepts from a model's parameters. Existing approaches typically rely on keywords to identify the target concept to be unlearned. However, we show that this keyword-based formulation is inherently limited: a visual concept is multi-dimensional, can be expressed in diverse textual forms, and often overlap with related concepts in the latent space, making keyword-only unlearning, which imprecisely indicate the target concept is brittle and prone to over-forgetting. This occurs because a single keyword represents only a narrow point estimate of the concept, failing to cover its full semantic distribution and entangled variations in the latent space. To address this limitation, we propose Diversified Unlearning, a distributional framework that represents a concept through a set of contextually diverse prompts rather than a single keyword. This richer representation enables more precise and robust unlearning. Through extensive experiments across multiple benchmarks and state-of-the-art baselines, we demonstrate that integrating Diversified Unlearning as an add-on component into existing unlearning pipelines consistently achieves stronger erasure, better retention of unrelated concepts, and improved robustness against adversarial recovery attacks.

Although text-to-image diffusion models exhibit remarkable generative power, concept erasure techniques are essential for their safe deployment to prevent the creation of harmful content. This has fostered a dynamic interplay between the development of erasure defenses and the adversarial probes designed to bypass them, and this co-evolution has progressively enhanced the efficacy of erasure methods. However, this adversarial co-evolution has converged on a narrow, text-centric paradigm that equates erasure with severing the text-to-image mapping, ignoring that the underlying visual knowledge related to undesired concepts still persist. To substantiate this claim, we investigate from a visual perspective, leveraging DDIM inversion to probe whether a generative pathway for the erased concept can still be found. However, identifying such a visual generative pathway is challenging because standard text-guided DDIM inversion is actively resisted by text-centric defenses within the erased model. To address this, we introduce TINA, a novel Text-free INversion Attack, which enforces this visual-only probe by operating under a null-text condition, thereby avoiding existing text-centric defenses. Moreover, TINA integrates an optimization procedure to overcome the accumulating approximation errors that arise when standard inversion operates without its usual textual guidance. Our experiments demonstrate that TINA regenerates erased concepts from models treated with state-of-the-art unlearning. The success of TINA proves that current methods merely obscure concepts, highlighting an urgent need for paradigms that operate directly on internal visual knowledge.

Recent progress in image generation models (IGMs) enables high-fidelity content creation but also amplifies risks, including the reproduction of copyrighted content and the generation of offensive content. Image Generation Model Unlearning (IGMU) mitigates these risks by removing harmful concepts without full retraining. Despite growing attention, the robustness under adversarial inputs, particularly image-side threats in black-box settings, remains underexplored. To bridge this gap, we present REFORGE, a black-box red-teaming framework that evaluates IGMU robustness via adversarial image prompts. REFORGE initializes stroke-based images and optimizes perturbations with a cross-attention-guided masking strategy that allocates noise to concept-relevant regions, balancing attack efficacy and visual fidelity. Extensive experiments across representative unlearning tasks and defenses demonstrate that REFORGE significantly improves attack success rate while achieving stronger semantic alignment and higher efficiency than involved baselines. These results expose persistent vulnerabilities in current IGMU methods and highlight the need for robustness-aware unlearning against multi-modal adversarial attacks. Our code is at: https://github.com/Imfatnoily/REFORGE.

Transformer based diffusion and vision-language models have achieved remarkable success; yet, efficiently removing undesirable or sensitive information without retraining remains a central challenge for model safety and compliance. We introduce Ratio-Aware Zero/One-step Optimized Retentive unlearning (RAZOR), a lightweight, model-agnostic unlearning framework that generalizes forgetting updates to coordinated multi-layer and multi-head edits within transformer backbones. RAZOR identifies the most important layers and attention heads by measuring how much they contribute to forgetting the target data while preserving useful knowledge. Then, it updates these parts of the model using a carefully regularized rule to avoid harming overall performance. The set of edited components grows gradually, ensuring precise unlearning without over-editing or damaging unrelated capabilities. We evaluate RAZOR on CLIP, Stable Diffusion, and vision-language models (VLMs) using widely adopted unlearning benchmarks covering identity, style, and object erasure tasks. Our results show that RAZOR achieves highly accurate and stable forgetting, even under quantization. This approach offers stronger retention and better efficiency than prior methods. Notably, it also operates significant faster than conventional techniques. These results demonstrate that RAZOR is a practical and scalable solution for safe, adaptive unlearning in transformer-based vision models.

Machine unlearning aims to remove specific outputs from trained models, often at the concept level, such as forgetting all occurrences of a particular celebrity or filtering content via text prompts. However, many undesired outputs, such as an individual's face or generations culturally or factually misinterpreted, cannot often be specified by text prompts. We address this underexplored setting of instance unlearning for outputs that are undesired but unpromptable, where the goal is to forget target outputs selectively while preserving the rest. To this end, we introduce an effective surrogate-based unlearning method that leverages image editing, timestep-aware weighting, and gradient surgery to guide trained diffusion models toward forgetting specific outputs. Experiments on conditional (Stable Diffusion 3) and unconditional (DDPM-CelebA) diffusion models demonstrate that our prompt-free method uniquely unlearns unpromptable outputs, such as faces and culturally inaccurate depictions, with preserved integrity, unlike prompt-based and prompt-free baselines. Our proposed method would serve as a practical hotfix for diffusion model providers to ensure privacy protection and ethical compliance.

Concept erasure is extensively utilized in image generation to prevent text-to-image models from generating undesired content. Existing methods can effectively erase narrow concepts that are specific and concrete, such as distinct intellectual properties (e.g. Pikachu) or recognizable characters (e.g. Elon Musk). However, their performance degrades on broad concepts such as sexual'' or violent'', whose wide scope and multi-faceted nature make them difficult to erase reliably. To overcome this limitation, we exploit the model's intrinsic embedding geometry to identify latent embeddings that encode a given concept. By clustering these embeddings, we derive a set of concept prototypes that summarize the model's internal representations of the concept, and employ them as negative conditioning signals during inference to achieve precise and reliable erasure. Extensive experiments across multiple benchmarks show that our approach achieves substantially more reliable removal of broad concepts while preserving overall image quality, marking a step towards safer and more controllable image generation.

The powerful generative capabilities of diffusion models have raised growing privacy and safety concerns regarding generating sensitive or undesired content. In response, machine unlearning (MU) -- commonly referred to as concept erasure (CE) in diffusion models -- has been introduced to remove specific knowledge from model parameters meanwhile preserving innocent knowledge. Despite recent advancements, existing unlearning methods often suffer from excessive and indiscriminate removal, which leads to substantial degradation in the quality of innocent generations. To preserve model utility, prior works rely on compensation, i.e., re-assimilating a subset of the remaining data or explicitly constraining the divergence from the pre-trained model on remaining concepts. However, we reveal that generations beyond the compensation scope still suffer, suggesting such post-remedial compensations are inherently insufficient for preserving the general utility of large-scale generative models. Therefore, in this paper, we advocate for developing compensation-free concept erasure operations, which precisely identify and eliminate the undesired knowledge such that the impact on other generations is minimal. In technique, we propose to MiM-MU, which is to unlearn a concept by minimizing the mutual information with a delicate design for computational effectiveness and for maintaining sampling distribution for other concepts. Extensive evaluations demonstrate that our proposed method achieves effective concept removal meanwhile maintaining high-quality generations for other concepts, and remarkably, without relying on any post-remedial compensation for the first time.

Removing undesired concepts from large-scale text-to-image (T2I) and text-to-video (T2V) diffusion models while preserving overall generative quality remains a major challenge, particularly as modern models such as Stable Diffusion v3, Flux, and OpenSora employ flow-matching and transformer-based architectures and extend to long-horizon video generation. Existing concept erasure methods, designed for earlier T2I/T2V models, often fail to generalize to these paradigms. To address this issue, we propose EraseAnything++, a unified framework for concept erasure in both image and video diffusion models with flow-matching objectives. Central to our approach is formulating concept erasure as a constrained multi-objective optimization problem that explicitly balances concept removal with preservation of generative utility. To solve the resulting conflicting objectives, we introduce an efficient utility-preserving unlearning strategy based on implicit gradient surgery. Furthermore, by integrating LoRA-based parameter tuning with attention-level regularization, our method anchors erasure on key visual representations and propagates it consistently across spatial and temporal dimensions. In the video setting, we further enhance consistency through an anchor-and-propagate mechanism that initializes erasure on reference frames and enforces it throughout subsequent transformer layers, thereby mitigating temporal drift. Extensive experiments on both image and video benchmarks demonstrate that EraseAnything++ substantially outperforms prior methods in erasure effectiveness, generative fidelity, and temporal consistency, establishing a new state of the art for concept erasure in next-generation diffusion models.

Deployed text-to-image diffusion models increasingly require post-hoc concept unlearning for copyright claims, artist opt-outs, safety updates, and protected-content mitigation without full retraining. A central challenge is erase-retain imbalance, aggressive updates suppress targets but damage shared capabilities, while conservative or anchor-based updates preserve quality yet leave concepts recoverable through related, compositional, paraphrased, or adversarial prompts. Inspired by retroactive interference, we propose SurgUn, which treats forgetting as controlled competition rather than direct deletion or one-to-one reassignment. SurgUn instantiates retroactive concept interference via distractor-conditioned gradient competition: target-gradient ascent weakens target-conditioned denoising or flow-matching behavior, while descent over a semantically diverse distractor set introduces competing non-target trajectories under the same prompt context. This redistributes outputs across multiple non-target modes instead of collapsing to a single proxy. To limit collateral forgetting through shared pathways, SurgUn adds pixel-grounded weight-space localization, a lightweight diagnostic that selects attention blocks by generated-image erase-retain behavior, exploiting the asymmetry that suppression is broadly achievable whereas retention is block-selective. Across UnlearnCanvas, IP-character erasure, Holistic Unlearning, EraseBench, and Ring-A-Bell on Stable Diffusion v1.5, SDXL, and SANA-1.5, SurgUn achieves a stronger erase-retain balance than baselines. Ablations show that diverse distractors, contrastive competition, and localization are all necessary for robust suppression while preserving related and unrelated concepts.

Pruning-based unlearning has recently emerged as a fast, training-free, and data-independent approach to remove undesired concepts from diffusion models. It promises high efficiency and robustness, offering an attractive alternative to traditional fine-tuning or editing-based unlearning. However, in this paper we uncover a hidden danger behind this promising paradigm. We find that the locations of pruned weights, typically set to zero during unlearning, can act as side-channel signals that leak critical information about the erased concepts. To verify this vulnerability, we design a novel attack framework capable of reviving erased concepts from pruned diffusion models in a fully data-free and training-free manner. Our experiments confirm that pruning-based unlearning is not inherently secure, as erased concepts can be effectively revived without any additional data or retraining. Extensive experiments on diffusion-based unlearning based on concept related weights lead to the conclusion: once the critical concept-related weights in diffusion models are identified, our method can effectively recover the original concept regardless of how the weights are manipulated. Finally, we explore potential defense strategies and advocate safer pruning mechanisms that conceal pruning locations while preserving unlearning effectiveness, providing practical insights for designing more secure pruning-based unlearning frameworks.

Recent advances in text-to-image (T2I) diffusion models have seen rapid and widespread adoption. However, their powerful generative capabilities raise concerns about potential misuse for synthesizing harmful, private, or copyrighted content. To mitigate such risks, concept erasure techniques have emerged as a promising solution. Prior works have primarily focused on fine-tuning the denoising component (e.g., the U-Net backbone). However, recent causal tracing studies suggest that visual attribute information is localized in the early self-attention layers of the text encoder, indicating a potential alternative for concept erasing. Building on this insight, we conduct preliminary experiments and find that directly fine-tuning early layers can suppress target concepts but often degrades the generation quality of non-target concepts. To overcome this limitation, we propose High-Level Representation Misdirection (HiRM), which misdirects high-level semantic representations of target concepts in the text encoder toward designated vectors such as random directions or semantically defined directions (e.g., supercategories), while updating only early layers that contain causal states of visual attributes. Our decoupling strategy enables precise concept removal with minimal impact on unrelated concepts, as demonstrated by strong results on UnlearnCanvas and NSFW benchmarks across diverse targets (e.g., objects, styles, nudity). HiRM also preserves generative utility at low training cost, transfers to state-of-the-art architectures such as Flux without additional training, and shows synergistic effects with denoiser-based concept erasing methods.

Text guided diffusion models are used by millions of users, but can be easily exploited to produce harmful content. Concept unlearning methods aim at reducing the models' likelihood of generating harmful content. Traditionally, this has been tackled at an individual concept level, with only a handful of recent works considering more realistic concept combinations. However, state of the art methods depend on full finetuning, which is computationally expensive. Concept localisation methods can facilitate selective finetuning, but existing techniques are static, resulting in suboptimal utility. In order to tackle these challenges, we propose TRUST (Targeted Robust Selective fine Tuning), a novel approach for dynamically estimating target concept neurons and unlearning them through selective finetuning, empowered by a Hessian based regularization. We show experimentally, against a number of SOTA baselines, that TRUST is robust against adversarial prompts, preserves generation quality to a significant degree, and is also significantly faster than the SOTA. Our method achieves unlearning of not only individual concepts but also combinations of concepts and conditional concepts, without any specific regularization.

Real-world deployment of text-to-image diffusion models requires continual concept removal as new privacy, copyright, or safety obligations arise over time. Existing unlearning methods, however, are designed for single-step deletion and collapse after only 3-5 sequential applications. We trace this instability to two compounding factors: (i) coarse mapping targets that cause degradation to accumulate unnecessarily across steps, and (ii) the absence of local protection for semantically neighboring concepts, whose shared internal representations make them the first to suffer collateral damage. Because this damage is strongest in the local semantic neighborhood of the forget concept, global replay alone cannot prevent it. Building on this analysis, we propose Locality-Aware Continual Unlearning (LACU), a framework with two complementary mechanisms. Locality-Aware Target Selection chooses, for each forget prompt, the context-preserving mapping prompt that the diffusion model itself treats as most similar to the original prompt, measured by score-prediction distance (how differently the model denoises the same noisy image under two text conditions), ensuring each update is as small and targeted as possible. Locality-Aware Replay uses the same metric to identify the retain concepts closest to the forget concept in the model's own representation and replays them as a local functional regularizer, directly shielding the most vulnerable neighborhood. Combined with teacher-student distillation and lightweight $\ell_2$ parameter regularization, LACU maintains stable unlearning over 10 sequential steps, preserving significantly higher related retention ($RR_{\text{acc}}$) and general retention ($GR_{\text{acc}}$) than recent baselines.

Concept erasure methods aim to remove specific unsafe target concepts in diffusion models while preserving image generation utility. To address the vulnerability that erased concepts can be easily recovered under adversarial attacks, adversarial concept erasure methods integrate adversarial optimization into the concept erasure process. However, existing adversarial concept erasure methods face a trade-off between robustness and computational cost. We attribute this to adversarial optimization techniques that use random samples to approximate the adversarial objective function. Adversarial optimization that uses a small number of samples fails to produce adversarial embeddings that accurately capture the target concept space. To mitigate this limitation, we propose Semantic-Guided Adversarial Optimization, which uses a single sample to produce adversarial embeddings that better capture the target concept space. We also propose Semantic-Guided Concept Erasure, which automatically maps the target concept to a semantically similar surrogate. Extensive experiments on not-safe-for-work content, artistic styles, and object-related concepts demonstrate that our method, S-GRACE (Semantic-Guided Robust Adversarial Concept Erasure) achieves state-of-the-art erasure robustness and superior image generation utility, with significantly lower computational cost than existing methods. Our code is available at https://github.com/Qhong-522/S-GRACE.

Data unlearning aims to remove the influence of specific training samples from a trained model. In fine-tuning methods, data unlearning relies primarily on loss maximization over forget samples, which often leads to quality degradation or incomplete forgetting. Existing methods perform unlearning uniformly across diffusion stages, ignoring diffusion dynamics from noise to data. Our systematic study of diffusion phases shows that forgetting in diffusion models is uneven across time and frequency, with theoretical justification of distributive distortion and forgetting-utility trade-off. By selectively forgetting time and frequency in diffusion models, we achieve both higher unlearning success rates and improved generation quality across diverse settings, including both conditional and unconditional scenarios. We also introduce an improved SSCD metric that measures dissimilarity using a normalized perturbation distance. Together, we provide practical insights for understanding and improving data unlearning in diffusion models.

As generative models become increasingly powerful and pervasive, the ability to unlearn specific data due to privacy concerns, legal requirements, or harmful content has become increasingly important. Unlearning aims to selectively remove the influence of particular data without costly retraining from scratch. Effective unlearning requires (i) forgetting undesired data, (ii) preserving generation quality, (iii) retaining the influence of desired training data on model parameters, and (iv) computational efficiency. We propose fast unlearning algorithms based on loss gradient orthogonalization for unconditional and conditional generative models. We consider the loss function of the original model restricted to the retain data and augment it by a term promoting the orthogonality between the two gradients of the loss function evaluated at the retain dataset and the forget dataset. We coin this method unlearning via orthogonalization (UNO) and show that the additional weak orthogonality constraint accelerates unlearning. The inclusion of the weak orthogonality constraint implies that each gradient descent step utilizes local information about the geometry via the Hessian of the loss function. Our geometry aware gradient descent step significantly accelerates the learning of the retain data loss function compared to a gradient descent which uses only gradient information. By alternating between an UNO and a gradient surgery step, we obtain UNO-S, which further reduces execution time. We demonstrate that our algorithms effectively forget data while preserving the fidelity of the original model. On standard image benchmarks, they achieve orders-of-magnitude faster unlearning than existing methods such as gradient surgery. We validate our approach on datasets of increasing complexity (MNIST, CelebA, and ImageNet-1K) and generative architectures ranging from VAEs to diffusion transformers.

Diffusion models excel at generating high-quality images but can memorize and reproduce harmful concepts when prompted. Although fine-tuning methods have been proposed to unlearn a target concept, they struggle to fully erase it while maintaining generation quality on other concepts, leaving models vulnerable to jailbreak attacks. Existing jailbreak methods demonstrate this vulnerability but offer limited insight into how unlearned models retain harmful concepts, limiting progress on effective defenses. In this work, we show that a coherent, interpretable, attack-accessible linear residual of the erased concept can be recovered in the token embedding space, and that both an attack and a defense follow from this structure. We introduce SubAttack, a novel jailbreaking attack that reads out this subspace by learning an orthogonal set of attack token embeddings, each being a linear combination of human-interpretable textual elements, revealing that unlearned models still retain the target concept through related textual components. Furthermore, our attack is also more powerful and transferable across text prompts, initial noises, and unlearned models than prior attacks. Conversely, projecting out the same subspace yields SubDefense, a lightweight plug-and-play defense mechanism that suppresses the residual concept in unlearned models. SubDefense provides stronger robustness than existing defenses while better preserving safe generation quality. Extensive experiments across multiple unlearning methods, concepts, and attack types demonstrate that our approach advances both understanding and mitigation of vulnerabilities in diffusion unlearning.