Diffusion Probabilistic Models (DPMs) have recently become prominent in generative modeling, showing excellent performance in producing high-fidelity, diverse data in various domains. Despite their remarkable generative capabilities, a significant dra...
Diffusion Probabilistic Models (DPMs) have recently become prominent in generative modeling, showing excellent performance in producing high-fidelity, diverse data in various domains. Despite their remarkable generative capabilities, a significant drawback of DPMs is their substantial computational cost and latency due to the inherently iterative, multi-step sampling process. This inefficiency poses a critical barrier to their practical deployment in real-world applications, motivating extensive research into reducing their inference time without compromising generation quality. This dissertation addresses the crucial challenge posed by efficiency-performance trade-offs of DPMs through novel methodologies across three distinct research directions. Our primary goal is to achieve better efficiency-performance relationship by reducing the number of function evaluations (NFEs) required for sample generation, while maintaining or improving sample quality.
First, we focus on improving the robustness of DPMs to varying NFEs by proposing a novel training objective. We introduce an observation-guided diffusion probabilistic model (OGDM) that reestablishes the training objective by integrating guidance from an observation process into the diffusion Markov chain in a principled way. This is achieved by introducing an additional loss term derived from the observation based on a conditional discriminator on noise level, which employs a Bernoulli distribution indicating whether its input lies on the (noisy) real manifold or not. This approach allows for more accurate log-likelihood optimization, especially under a limited NFE budget during inference, leading to better denoising networks that maintain quality even with faster sampling, without incurring extra inference costs.
Second, we delve into enhancing existing ODE-based DPM solvers through the extrapolation sampling algorithm. Inspired by Richardson extrapolation, we propose RX-DPM, a method that leverages an additional ODE solution at intermediate time steps to extrapolate denoised predictions. This significantly improves the numerical accuracy of the sampling process, reducing truncation errors and leading to higher-quality samples for a given NFE. Notably, RX-DPM is generalized for arbitrary time step scheduling and can be seamlessly integrated into various existing solvers with minimal computational overhead. Additionally, RX-DPM provides explicit error estimates, effectively demonstrating the faster convergence due to the higher order of the leading error term.
Third, we investigate knowledge distillation techniques to create one-step student generators from multi-step teacher DPMs. Recent advancements in this area have often incorporated adversarial objectives, leveraging GAN training to guide the student towards higher fidelity and faster convergence. Building upon this approach, we formulate a principled objective that forces the distributions of the student and teacher models to be identical within the discriminator's feature space. We utilize characteristic functions---which uniquely determine probability distributions---adopting a squared CFD metric approximated via their empirical counterparts. This batch-level statistical comparison provides a holistic distributional alignment, enhancing adversarial training and exploring novel optimization pathways.
The contributions of the proposed training, sampling and distillation frameworks are validated through extensive experiments and analyses across standard benchmarks, respectively, demonstrating the effectiveness in improving the efficiency-performance trade-offs of DPMs.