Efficient training and sampling of diffusion-based generative models
Abstract
The recent surge of AI generative models has led to significant breakthroughs in content creation, particularly with diffusion-based models such as diffusion models and rectified flow models (a.k.a. flow matching models). However, these models suffer from limitations in training efficiency, sampling speed, gradient consistency, and others. This thesis proposes a suite of three methods that tries to address these challenges from complementary perspectives. First, we improved the training efficiency of diffusion-based models. Recent studies have shown that different diffusion methods and architectures trained on the same dataset tend to produce similar results for the same input noise. This observation suggests that certain noise inputs consistently guide the models to generate a specific sample. Those noise inputs are referred to as ``preferable noise'' for that sample. By visualizing the noise-sample pairs of rectified flow models and stable diffusion models in two-dimensional spaces, it is observed that the preferable paths, connecting preferable noises to the corresponding samples, are better organized with significantly fewer crossings compared with the random paths, connecting random noises to training samples. The path crossings in two-dimensional spaces indicate a shorter inter-path distance in the corresponding high-dimensional spaces. Inspired by this observation, this study proposes the Distance-Aware Noise-Sample Matching (DANSM) method to lengthen the inter-path…
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.