Towards smarter, safer, and swifter generative AI: adapting, guarding, and accelerating diffusion models
Abstract
Diffusion models (DMs) have established a new state-of-the-art in generative artificial intelligence, demonstrating unparalleled success in synthesizing high-fidelity images, videos, and continuous sequences. However, their widespread real-world deployment is hindered by three fundamental limitations: rigid Gaussian assumptions that fail to accommodate non-standard data types, the risk of generating inappropriate or sensitive content, and the high computational cost of iterative inference. This dissertation addresses these challenges by developing a suite of specialized methodologies designed to build the next generation of Smarter, Safer, and Swifter generative models. To achieve smarter generation, we first generalize the theoretical foundation of DMs through the lens of Bregman divergence, extending their applicability to count and range-bounded data domains. We further enhance model intelligence by introducing a dense reward perspective for preference alignment, prioritizing the formative early steps of the diffusion trajectory to better follow complex human instructions. For safer generation, we introduce Score Forgetting Distillation (SFD), a swift, data-free machine unlearning method that fosters trustworthiness by erasing unsafe concepts through efficient fine-tuning rather than prohibitive re-training. Finally, to make models swifter, we present DLM-One, a score-distillation-based framework that enables near-instantaneous one-step sequence generation in continuous…
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.