Jiakui/MV-AR

Auto-Regressively Generating Multi-View Consistent Images

JiaKui Hu*, Yuxiao Yang*, Jialun Liu, Jinbo Wu, Chen Zhao, Yanye Lu
PKU, BaiduVis, THU

Introduction

Diffusion-based multi-view image generation methods use a specific reference view for predicting subsequent views, which becomes problematic when overlap between the reference view and the predicted view is minimal, affecting image quality and multi-view consistency. Our MV-AR addresses this by using the preceding view with significant overlap for conditioning.

Results

Text to Multiview images

Image to Multiview images

Text + Geometric to Multiview images

Quick Start

Requirements

Please follow the instructions in code.

Reproduce

1. Please download flan-t5-xl in ./pretrained_models;
2. Please download Cap3D_automated_Objaverse_full.csv in dataset/captions;
3. Please download models here, put them in ./pretrained_models;
4. Run:

# For t2mv on objaverse
sh sample_tcam2i.sh
# For t2mv on GSO
sh sample_icam2i_gso.sh
# For i2mv on GSO
sh sample_icam2i_gso.sh

The generated images will be saved to samples_objaverse_nv_ray/.

Acknowledgement

This repository is heavily based on LlamaGen. We would like to thank the authors of these work for publicly releasing their code.

For help or issues using this git, please feel free to submit a GitHub issue.

For other communications related to this git, please contact jkhu29@stu.pku.edu.cn.

Citation

@article{hu2025mvar,
  title={Auto-Regressively Generating Multi-View Consistent Images},
  author={Hu, JiaKui and Yang, Yuxiao and Liu, Jialun and Wu, Jinbo and Zhao, Chen and Lu, Yanye},
  journal={arXiv preprint arXiv:2506.18527},
  year={2025}
}