Back to search

McMaster University

Explicit Motion Representations for Generative Visual Models

Abstract

dc:description.abstract

Generative visual models have achieved strong performance in image and video synthesis, but precise control over motion remains difficult. This thesis studies explicit point-based motion representations for generative visual models. The central argument is that motion can serve as a structured bridge between visual content and generative synthesis when it is represented explicitly, anchored to meaningful references, and stabilized during generation. The thesis develops this argument in two complementary settings. The first setting is sparse motion control for image editing. GoodDrag uses externally specified handle-target point pairs to guide diffusion-based image editing. It addresses two failure modes of diffusion drag editing: accumulated latent perturbation and handle-point feature drift. Alternating Drag and Denoising interleaves motion supervision with diffusion denoising to preserve image fidelity, while Information-Preserving Motion Supervision anchors motion control to the original handle-point feature to preserve semantic correspondence. The second setting is dense future motion generation from video history. TrajLoom generates dense future trajectories and visibility from observed video and past motion. It represents trajectories using Grid-Anchor Offset Encoding, compresses dense motion into coherent latent tokens with TrajLoom-VAE, and samples future motion with TrajLoom-Flow, a history-conditioned flow-matching generator. Boundary hints and on-policy fine-tuning improve history-future continuity and reduce long-horizon drift. Together, GoodDrag and TrajLoom study opposite directions of the relation between motion and generation: using given motion to control visual editing, and generating motion as an intermediate representation for future video synthesis. The thesis shows that explicit motion representations can improve controllability, temporal coherence, and interpretability in generative visual models.

Degree

thesis:*
Department dc:contributor.department
Electrical and Computer Engineering
Year dc:date.issued
2026

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Zhang, Zewei
Advisor dc:contributor.advisor
  • Chen, Jun

Subjects

dc:subject × 10

Rights

dc:rights
Statement dc:rights
  • Attribution-NonCommercial-NoDerivatives 4.0 International
Language dc:language.iso
en

Identifiers

dc:identifier.*

Chain of custody

source
Harvested from
McMaster University
Base URL
macsphere.mcmaster.ca/server/oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Zhang, Zewei. Explicit Motion Representations for Generative Visual Models. 2026. https://hdl.handle.net/11375/33607