Access Type

Open Access Dissertation

Date of Award

January 2025

Degree Type

Dissertation

Degree Name

Ph.D.

Department

Computer Science

First Advisor

Jing Hua

Abstract

Generative Artificial Intelligence (AI) has transformed how we synthesize, edit, and manipulate complex data, opening new possibilities for immersive digital content. This dissertation, titled “Advancing Generative AI in 3D and 4D Spaces,” investigates the capabilities and limitations of state-of-the-art generative models in static 3D scenes and time-aware 4D environments. In the 3D setting, scene editing is hindered by multi-view inconsistency and the high computational cost of per-scene retraining. To address these issues, the work introduces Free-Editor, a training free approach that utilizes an Edit Transformer to propagate a single edited view across all camera perspectives without additional optimisation, delivering prompt faithful edits up to twenty times faster than contemporary baselines while preserving geometric fidelity. Extending into the 4D domain, the dissertation confronts the added complexity of temporal coherence. A new framework, PSF-4D, adapts diffusion-based generation to dynamic scenes through progressive, correlated noise sampling that couples spatial and temporal information, enabling coherent local edits, style transfers, and object removals throughout video-volumes. Together, these contributions push generative AI beyond static content creation toward real-time, spatiotemporally consistent scene manipulation, with broad implications for virtual production, robotics simulation, and immersive analytics.

Share

COinS