Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

VideoGPA: Distilling Geometry Priors for 3D-Consistent Video Generation

Hongyang Du, Junjie Ye, Xiaoyan Cong, Runhao Li, Jingcheng Ni, Aman Agarwal, Zeqi Zhou, Zekun Li, Randall Balestriero, Yue Wang

TL;DR

VideoGPA tackles 3D inconsistency in pretrained video diffusion models by leveraging geometry foundation models (GFMs) to derive dense 3D signals and a reprojection-based 3D consistency score. This self-supervised signal is used to construct preference pairs, which guide a $v$-prediction diffusion model through Direct Preference Optimization, with LoRA-based fine-tuning involving only about $1\%$ of parameters and roughly $2{,}500$ preference pairs. The method improves 3D reconstruction fidelity, cross-view consistency, and human-perceived quality in both image-to-video and text-to-video tasks, outperforming SFT, Epipolar-DPO, and GeoVideo baselines. By enforcing scene-level geometric coherence, VideoGPA also acts as a geometric regularizer for motion generation, improving temporal stability and physical plausibility with minimal extra training.

Abstract

While recent video diffusion models (VDMs) produce visually impressive results, they fundamentally struggle to maintain 3D structural consistency, often resulting in object deformation or spatial drift. We hypothesize that these failures arise because standard denoising objectives lack explicit incentives for geometric coherence. To address this, we introduce VideoGPA (Video Geometric Preference Alignment), a data-efficient self-supervised framework that leverages a geometry foundation model to automatically derive dense preference signals that guide VDMs via Direct Preference Optimization (DPO). This approach effectively steers the generative distribution toward inherent 3D consistency without requiring human annotations. VideoGPA significantly enhances temporal stability, physical plausibility, and motion coherence using minimal preference pairs, consistently outperforming state-of-the-art baselines in extensive experiments.

VideoGPA: Distilling Geometry Priors for 3D-Consistent Video Generation

TL;DR

VideoGPA tackles 3D inconsistency in pretrained video diffusion models by leveraging geometry foundation models (GFMs) to derive dense 3D signals and a reprojection-based 3D consistency score. This self-supervised signal is used to construct preference pairs, which guide a -prediction diffusion model through Direct Preference Optimization, with LoRA-based fine-tuning involving only about of parameters and roughly preference pairs. The method improves 3D reconstruction fidelity, cross-view consistency, and human-perceived quality in both image-to-video and text-to-video tasks, outperforming SFT, Epipolar-DPO, and GeoVideo baselines. By enforcing scene-level geometric coherence, VideoGPA also acts as a geometric regularizer for motion generation, improving temporal stability and physical plausibility with minimal extra training.

Abstract

While recent video diffusion models (VDMs) produce visually impressive results, they fundamentally struggle to maintain 3D structural consistency, often resulting in object deformation or spatial drift. We hypothesize that these failures arise because standard denoising objectives lack explicit incentives for geometric coherence. To address this, we introduce VideoGPA (Video Geometric Preference Alignment), a data-efficient self-supervised framework that leverages a geometry foundation model to automatically derive dense preference signals that guide VDMs via Direct Preference Optimization (DPO). This approach effectively steers the generative distribution toward inherent 3D consistency without requiring human annotations. VideoGPA significantly enhances temporal stability, physical plausibility, and motion coherence using minimal preference pairs, consistently outperforming state-of-the-art baselines in extensive experiments.
Paper Structure (44 sections, 12 equations, 7 figures, 6 tables, 1 algorithm)

This paper contains 44 sections, 12 equations, 7 figures, 6 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Works
  3. Video Generation Models
  4. Video Diffusion Alignment
  5. Geometric Foundation Models
  6. Methodology
  7. Preliminaries
  8. Direct Preference Optimization
  9. From Policy to Diffusion Log-likelihood
  10. DPO for $v$-Prediction Video Diffusion
  11. Preference Modeling
  12. 3D Consistency Score
  13. Post-Training Alignment
  14. Image-to-Video (I2V)
  15. Text-to-Video (T2V)
  16. ...and 29 more sections

Figures (7)

  • Figure 1: Overview of VideoGPA and representative results. (a) VideoGPA aligns a pretrained video diffusion model through proposed reconstruction-guided preference optimization. (b) Image-to-video examples comparing the base model yang2025cogvideoxtexttovideodiffusionmodels and VideoGPA, showing improved geometric stability under camera motion. (c) Text-to-video examples demonstrating improved structural coherence and reduced geometric artifacts.
  • Figure 2: Pipeline of VideoGPA. A geometric foundation model probes generated videos to assess scene-level 3D consistency, which is used to form self-supervised preference pairs for post-training alignment via DPO.
  • Figure 3: Human preference study on I2V generation. VideoGPA is most frequently preferred, indicating improved perceptual quality and 3D consistency.
  • Figure 4: Qualitative comparison on I2V generation. We compare VideoGPA with the base model, SFT, and Epipolar-DPO. Highlighted regions illustrate improvements in (a) structural and geometric consistency, (b) texture stability and de-flickering, (c) robustness under challenging lighting, and (d) object attribute consistency.
  • Figure 5: Scene-level v.s. local geometry. Comparison between local geometric metrics and the proposed scene-level 3D consistency score on a corrupted video. Local, pairwise constraints yield a false positive, while the scene-level metric correctly identifies geometric inconsistency.
  • ...and 2 more figures