Inference-time editing and guidance methods using diffusion-based generative models

This thesis explores the efficient and effective editing and guidance algorithms of diffusion-based generative models across three distinct tasks: prompt-based video editing, offline model-based design optimization and conditional image generation. First, we propose FastVideoEdit, an efficient zero-shot text-to-video editing method leveraging consistency models to bypass computational bottlenecks of typical editing pipelines. By directly mapping source videos to edited output, while preserving visual fidelity and temporal coherence, significantly reducing running time. Second, we present DEMO, a two-phase editing approach for offline model-based optimization. DEMO first generates high-performing candidates via gradient ascent on a surrogate model, then calibrates these designs with a learned diffusion prior to overcome out-of-distribution (OOD) issues and ensure realism and distributional validity of generated samples. Lastly, we introduce ADMMDiff, a novel training-free conditional diffusion framework that decouples unconditional generation and guidance via the Alternating Direction Method of Multipliers (ADMM), achieving superior performance across multiple tasks including image generation with nonlinear guidance, linear inverse problems, and controllable motion synthesis. Together, these works demonstrate the versatility and effectiveness of diffusion models in enabling efficient editing, trustworthy optimization, and controllable generation, advancing the frontiers of generative modeling in computer vision and optimization

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.