A Survey of Recent 3D Scene Generation

PhD Qualifying Examination


Title: "A Survey of Recent 3D Scene Generation"

by

Mr. Sizhe SONG


Abstract:

3D scene generation aims to create spatially coherent, semantically
meaningful, and photorealistic environments for uses ranging from immersive
media and robotics to autonomous driving and embodied AI. Earlier approaches
built on procedural rules were scalable but lacked variety. With the rise of
deep generative models (such as GANs and diffusion models) and modern 3D
representations (including NeRF and 3D Gaussians), it has become possible to
learn distributions of real-world scenes, resulting in better realism,
diversity, and multi-view consistency. Diffusion-based methods in particular
have recently strengthened the link between 3D scene synthesis and
photorealism by framing the task through image or video generation. This
survey reviews current progress in the field and groups existing techniques
into four main categories: procedural methods, neural 3D-centric
approaches, image-driven generation, and video-driven generation. We examine
their underlying principles, key trade-offs, and representative
achievements, and summarize widely used datasets, evaluation practices, and
practical applications. The survey concludes with a discussion of open
challenges related to model capacity, 3D representations, data and annotation
needs, and evaluation standards, while pointing to future directions such as
higher fidelity, physics-aware and interactive scene creation, and integrated
perception—generation systems.


Date:                   Tuesday, 27 January 2026

Time:                   2:00pm - 4:00pm

Venue:                  Room 3494
                        Lift 25/26

Committee Members:      Prof. Gary Chan (Supervisor)
                        Prof. Nevin Zhang (Chairperson)
                        Prof. Raymond Wong