Advancing Guidance Methods for Diverse, High-Quality, and Efficient Generation with Diffusion Models
Abstract
While diffusion models have recently emerged as the state-of-the-art approach for visual content creation, such as image and video synthesis, unguided sampling from these systems often results in low-quality samples that poorly align with the input condition. To address this, classifier-free guidance (CFG) has become an essential component in modern diffusion models for improving both generation quality and alignment with the conditioning signal. However, CFG typically requires relatively high guidance scales to be effective, and such high scales come with notable drawbacks. First, higher CFG scales significantly reduce generation diversity by restricting the model outputs to strong modes of the data distribution. Second, by overemphasizing the conditional prediction at each sampling step, CFG-guided generations often exhibit oversaturation and artifacts, especially at high guidance scales. CFG also doubles the sampling cost of diffusion models, as it requires two forward passes at each inference step. Combined with the iterative nature of the diffusion sampling process, this makes sampling inefficient and time-consuming. Finally, CFG requires a specific training procedure that learns the unconditional model alongside the conditional score; as a result, it is limited to conditional generation. In this thesis, our goal is to study and overcome the limitations of classifier-free guidance outlined above. We first show that overreliance on the conditioning signal is the main…
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.