Analysis of errors in generative image and video models

Image and video generative models have become increasingly powerful and produce visuals that are more difficult to distinguish from real ones in recent years. Users can create images and videos to make their imaginations come true easier than ever before. However, on closer examination, these models make interesting mistakes. In this thesis, we first discuss a systematic way of analyzing errors in generated images relating to projective geometry and shadows at a population level. We find that generated images can be reliably distinguished from real images by derived geometric features alone without looking at pixels. Then, we introduce methods that can effectively judge whether a generated video is physically plausible for robotic demonstrations.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Electrical & Computer Engr

Grantor

University of Illinois Urbana-Champaign

Year dc:date

2025

Author dc:creator

• Mai, Hanlin

Contributors dc:contributor

• Lazebnik, Svetlana

• Generative AI
• Diffusion Models

Statement dc:rights

• Copyright 2025 Hanlin Mai

Language dc:language

en

Handle dc:identifier

https://hdl.handle.net/2142/132611

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/132611