Introduces a layered intervention framework for knowledge infusion in multimodal generative models and empirically demonstrates complementarity of layers in a safety-alignment task with diffusion models.
hub
Red-teaming the stable diffusion safety filter
23 Pith papers cite this work. Polarity classification is still indexing.
23
Pith papers citing it
hub tools
representative citing papers
DivIn samples initial noise from a guidance potential posterior via Langevin dynamics to improve diversity in class-to-image and text-to-image generation.
Orthogonal Negative Guidance subtracts only the orthogonal component of negative-prompt attention features from positive ones in FLUX models to suppress concepts while preserving semantics and quality.
FlowErase-RL applies GRPO to reformulate concept erasure in flow matching models as reward optimization using a dynamic dual-path mechanism for target suppression and non-target preservation.
Mosaic combines text perturbation, multi-view image optimization, and surrogate model ensembles to reduce reliance on any single open-source model and achieve higher attack success rates on commercial closed-source VLMs.
HyPE detects harmful prompts as outliers in hyperbolic space and HyPS sanitizes them using explainable attribution, outperforming prior defenses in accuracy and robustness across datasets and adversarial scenarios.
A transcoder-based in-place replacement of the bottleneck layer enables selective concept removal in modern diffusion and autoregressive image models without degrading output quality.
REINS uses supervised PCA on safety-labeled activations to find a linear direction that, when added to hidden states at roughly 50% depth in video diffusion transformers, redirects generations from unsafe to safe content across multiple models.
UVR is a training-free framework that uses attention modulation based on identified information flow stages in multimodal DiT attention to erase unsafe semantics in image synthesis and editing at 91% and 77% rates while preserving quality.
RedEdit finds that fewer than two photo edits on average let 76.2% of unsafe images evade detectors while retaining 93.0% of malicious semantics.
DSR decomposes harmful intents into benign textual and visual primitives that MLLMs fuse into harmful outputs, achieving high attack success with low input toxicity.
BEAP is a black-box embedding-aware prompting attack using LLM-guided search that raises attack success rate over 60% against unlearned diffusion models while keeping prompts undetectable.
SafeDiffusion-R1 uses online GRPO with CLIP embedding steering to cut inappropriate content from 48.9% to 18.07% and nudity detections from 646 to 15 in diffusion models while raising GenEval scores from 42.08% to 47.83% and generalizing across seven harm categories without supervised pairs or extra
Gaussian probing infers harmful model specialization from parameter perturbations and internal representation responses to Gaussian latent ensembles rather than from generated outputs.
A training-free double-projection linear transformation erases target concepts from generative models by computing a proxy projection then applying a constrained update in the left null space of known directions.
Unlearning methods that strongly erase concepts from text-to-image diffusion models consistently degrade performance on attribute binding, spatial reasoning, and counting tasks.
VLBiasBench is a new large-scale benchmark with 128,342 samples covering nine social bias categories plus two intersectional ones to evaluate biases in LVLMs.
SalUn uses gradient-based weight saliency to achieve effective machine unlearning of data, classes, or concepts in image classification and generation, narrowing the gap to exact retraining.
Iterative self-improving codebooks enhance safety in autoregressive multimodal models by self-identifying unsafe generations and updating the codebook to eliminate harmful visual token mappings without external feedback.
Defines CARE score and proposes ReCARE framework to preserve co-occurring benign concepts during targeted unlearning in diffusion models.
DDiffusion uses semantic retrieval on prompt embeddings and localized editing inside the diffusion process to suppress NSFW content while avoiding binary allow/block signals.
SHIFT learns and applies steering vectors to selected layers and timesteps in DiT models to suppress concepts, shift styles, or bias objects while keeping image quality and prompt adherence intact.
citing papers explorer
-
Where Should Knowledge Enter? A Layered Framework for Knowledge Infusion in Multimodal Iterative Generative Model
Introduces a layered intervention framework for knowledge infusion in multimodal generative models and empirically demonstrates complementarity of layers in a safety-alignment task with diffusion models.
-
Initialization is Half the Battle: Generating Diverse Images from a Guidance Potential Posterior
DivIn samples initial noise from a guidance potential posterior via Langevin dynamics to improve diversity in class-to-image and text-to-image generation.
-
Orthogonal Negative Guidance in Attention Feature Space for Text-to-Image Generation
Orthogonal Negative Guidance subtracts only the orthogonal component of negative-prompt attention features from positive ones in FLUX models to suppress concepts while preserving semantics and quality.
-
FlowErase-RL: Rethinking Concept Erasure as Reward Optimization in Flow Matching Models
FlowErase-RL applies GRPO to reformulate concept erasure in flow matching models as reward optimization using a dynamic dual-path mechanism for target suppression and non-target preservation.
-
Mosaic: Multimodal Jailbreak against Closed-Source VLMs via Multi-View Ensemble Optimization
Mosaic combines text perturbation, multi-view image optimization, and surrogate model ensembles to reduce reliance on any single open-source model and achieve higher attack success rates on commercial closed-source VLMs.
-
Harnessing Hyperbolic Geometry for Harmful Prompt Detection and Sanitization
HyPE detects harmful prompts as outliers in hyperbolic space and HyPS sanitizes them using explainable attribution, outperforming prior defenses in accuracy and robustness across datasets and adversarial scenarios.
-
Concept Removal for Frontier Image Generative Models
A transcoder-based in-place replacement of the bottleneck layer enables selective concept removal in modern diffusion and autoregressive image models without degrading output quality.
-
Pulling The REINS: Training-Free Safety Alignment of Video Diffusion Models via Representation Steering
REINS uses supervised PCA on safety-labeled activations to find a linear direction that, when added to hidden states at roughly 50% depth in video diffusion transformers, redirects generations from unsafe to safe content across multiple models.
-
Unified Safe In-context Image Generation in Multimodal Diffusion Transformers via Restricting Unsafe Information Flows
UVR is a training-free framework that uses attention modulation based on identified information flow stages in multimodal DiT attention to erase unsafe semantics in image synthesis and editing at 91% and 77% rates while preserving quality.
-
RedEdit: Agentic Red-Teaming of Image Safety Classifiers via MCTS-Guided Photo-Editing
RedEdit finds that fewer than two photo edits on average let 76.2% of unsafe images evade detectors while retaining 93.0% of malicious semantics.
-
Benign Inputs, Harmful Outputs: Cross-Modal Jailbreaking via Distributed Semantic Recomposition
DSR decomposes harmful intents into benign textual and visual primitives that MLLMs fuse into harmful outputs, achieving high attack success with low input toxicity.
-
Erased but Exploitable: Black-box Embedding-Aware Prompting Against Unlearned Text-to-Image Diffusion Models
BEAP is a black-box embedding-aware prompting attack using LLM-guided search that raises attack success rate over 60% against unlearned diffusion models while keeping prompts undetectable.
-
SafeDiffusion-R1: Online Reward Steering for Safe Diffusion Post-Training
SafeDiffusion-R1 uses online GRPO with CLIP embedding steering to cut inappropriate content from 48.9% to 18.07% and nudity detections from 646 to 15 in diffusion models while raising GenEval scores from 42.08% to 47.83% and generalizing across seven harm categories without supervised pairs or extra
-
Evaluation without Generation: Non-Generative Assessment of Harmful Model Specialization with Applications to CSAM
Gaussian probing infers harmful model specialization from parameter perturbations and internal representation responses to Gaussian latent ensembles rather than from generated outputs.
-
Closed-Form Concept Erasure via Double Projections
A training-free double-projection linear transformation erases target concepts from generative models by computing a proxy projection then applying a constrained update in the left null space of known directions.
-
Erasure or Erosion? Evaluating Compositional Degradation in Unlearned Text-To-Image Diffusion Models
Unlearning methods that strongly erase concepts from text-to-image diffusion models consistently degrade performance on attribute binding, spatial reasoning, and counting tasks.
-
VLBiasBench: A Comprehensive Benchmark for Evaluating Bias in Large Vision-Language Model
VLBiasBench is a new large-scale benchmark with 128,342 samples covering nine social bias categories plus two intersectional ones to evaluate biases in LVLMs.
-
SalUn: Empowering Machine Unlearning via Gradient-based Weight Saliency in Both Image Classification and Generation
SalUn uses gradient-based weight saliency to achieve effective machine unlearning of data, classes, or concepts in image classification and generation, narrowing the gap to exact retraining.
-
Safe Autoregressive Image Generation with Iterative Self-Improving Codebooks
Iterative self-improving codebooks enhance safety in autoregressive multimodal models by self-identifying unsafe generations and updating the codebook to eliminate harmful visual token mappings without external feedback.
-
Co-occurring associated retained concepts in Diffusion Unlearning
Defines CARE score and proposes ReCARE framework to preserve co-occurring benign concepts during targeted unlearning in diffusion models.
-
Disciplined Diffusion: Text-to-Image Diffusion Model against NSFW Generation
DDiffusion uses semantic retrieval on prompt embeddings and localized editing inside the diffusion process to suppress NSFW content while avoiding binary allow/block signals.
-
SHIFT: Steering Hidden Intermediates in Flow Transformers
SHIFT learns and applies steering vectors to selected layers and timesteps in DiT models to suppress concepts, shift styles, or bias objects while keeping image quality and prompt adherence intact.
- GEM: Geometric Erasure by Contrastive Velocity Matching in Rectified Flows