Video Geometry Corrector (VGC): Fixing Geometric Inconsistency in Generated Videos

Forum topic · 小凯 · 2026-10-20

Summary

A forum post introduces a new arXiv paper (arXiv:2610.23856) presenting the Video Geometry Corrector (VGC), a post-hoc video-to-video model that improves geometric consistency in AI-generated videos. The authors observe that generated videos warp under viewpoint change, show inconsistent parallax, and sometimes collapse smooth camera motion into shot-like transitions. They attribute this to a data problem: web-scale video corpora mostly contain dynamic objects filmed by static or weakly moving cameras, providing sparse supervision for viewpoint-induced appearance changes. VGC is trained with synthetic paired supervision (VG-Pairs): real video clips with diverse camera motion and pose annotations are paired with distorted versions synthesized by a pretrained image-to-video inpainting model using trajectory-adaptive keyframes. VGC learns to map distorted inputs back to geometry-consistent targets without modifying the source generator, using optical flow as an auxiliary cue to disentangle camera and object motion. Experiments on pose-conditioned and text-driven benchmarks show improved geometric consistency across multiple source generators, and scaling to larger backbones extends the framework's general video refinement capability.

Paper Overview

Field: Computer Vision
Authors: Weijie Lyu, Tiancheng Shen, Jason Li, Yujing Wang, Xiaoyuan Wang, Haobo Yuan, Yizhou Zhao, Ming-Hsuan Yang
Published: 2026-10-16
arXiv: 2610.23856

Introduction

Generated videos often lose geometric consistency under camera motion: structures warp under viewpoint change, parallax becomes inconsistent, and smooth camera movement may collapse into shot-like transitions. The authors argue this failure is largely a data problem — in web-scale video corpora, most visible motion comes from dynamic objects captured by static or weakly moving cameras, leaving only sparse supervision for viewpoint-induced appearance changes.

Method: Video Geometry Corrector (VGC)

VGC is a post-hoc video-to-video model trained on synthetic paired supervision:

Starting from real videos with diverse camera motion and pose annotations, the authors construct VG-Pairs via a synthetic paired-supervision data generation pipeline.
Each video clip serves as a geometry-consistent target and is paired with a distorted version synthesized by a pretrained image-to-video inpainting model, using trajectory-adaptive keyframes.
VGC learns to map distorted inputs back to geometry-consistent targets without modifying the source generator.
Optical flow is used as an auxiliary cue to help disentangle camera motion from object motion.

Results

Experiments on pose-conditioned and text-driven benchmarks show that VGC improves geometric consistency across multiple source generators. Scaling to larger backbones further extends the framework's general video refinement capability.