Investigating Safety Compliance for Image Generation Models
Abstract
Text-to-Image (T2I) diffusion models have advanced rapidly, yet safe deployment remains challenging. T2I safety is multi-dimensional, involving bias, privacy, copyright, robustness, harmful-content control, and structural reliability. This thesis focuses on two urgent problems: anatomical reliability in human images and suppression of unsafe content. These dimensions are complementary: a model may avoid harmful concepts while still producing invalid human structures, or generate plausible humans while remaining unsafe. First, we address anatomical failures, especially distorted or incomplete hands under complex poses and occlusions. Such defects reduce user trust and limit downstream use. We propose SGMHand, a structure-guided inpainting framework with a Structure-Guided Modulation module and a Keypoint-Aware loss. By injecting skeletal and depth priors into denoising, SGMHand guides generation toward anatomically valid topology. Experiments across diffusion backbones show that it reduces fused fingers and implausible poses while preserving consistency and prompt faithfulness. Second, we study risk control for pornographic and violent imagery. Existing concept-erasure methods often damage benign semantics, causing over-erasure and quality loss. We propose OrthoEraser, which formulates erasure as geometric projection. Sparse Autoencoders disentangle latent activations and identify sensitive neurons, after which intervention directions are projected into the null space of…
Read the paper
Related studies
- Learning from First-Generation Qualitative Approaches in the IS Discipline: An Evolutionary View and Some Implications for Authors and Evaluators (PART 1/2)
- Commentary on the clinical and preclinical dosage limits of interstitially administered magnetic fluids for therapeutic hyperthermia based on current practice and efficacy models
- Prosthesis: New Technological Opportunities and Innovative Biomedical Devices
- The history and future of neural modeling for cochlear implants
- Ryanodine receptors: waking up from refractoriness
- Novel Transgenic Rabbit Model Sheds Light on the Puzzling Role of Matrix Metalloproteinase-12 in Atherosclerosis
Go deeper in Evidence Hub
Evidence Hub adds a study data sheet (design, sample size, evidence level, key numbers), an AI summary and a chat grounded in the paper's text.