ReImagine: Controllable High-Quality Human Video Generation via an Image-First Approach

Forum topic · 小凯 · 2026-04-23

Summary

ReImagine is a computer vision research paper (arXiv: 2604.19720) that revisits human video generation from an image-first perspective. The core idea is to decouple appearance modeling from temporal consistency: high-quality human appearance is learned through image generation and then used as a prior for video synthesis. The authors—Zhengwentai Sun, Keru Zheng, Chenghong Li, and colleagues—propose a pose- and viewpoint-controllable pipeline that combines a pretrained image backbone with SMPL-X-based motion guidance, followed by a training-free temporal refinement stage built on a pretrained video diffusion model. This design addresses the long-standing difficulty of jointly modeling human appearance, motion, and camera viewpoint under limited multi-view data, where existing methods often handle factors separately at the cost of controllability or visual quality. The method generates high-quality, temporally consistent videos across diverse poses and viewpoints. The team also releases a standardized human dataset and auxiliary models for compositional human image synthesis. Code and data are available at https://github.com/Taited/ReImagine.

Paper Overview

Field: Computer Vision (CV) Authors: Zhengwentai Sun, Keru Zheng, Chenghong Li, Hongjie Liao, Xihe Yang, Heyuan Li, Yihao Zhi, Shuliang Ning, Shuguang Cui, Xiaoguang Han Published: 2026-04-21 arXiv: 2604.19720

Abstract

Human video generation remains challenging due to the difficulty of jointly modeling human appearance, motion, and camera viewpoint under limited multi-view data. Existing methods often address these factors separately, resulting in limited controllability or reduced visual quality. The authors revisit this problem from an image-first perspective, where high-quality human appearance is learned via image generation and used as a prior for video synthesis, decoupling appearance modeling from temporal consistency.

Method Highlights

Pose- and viewpoint-controllable pipeline: combines a pretrained image backbone with SMPL-X-based motion guidance
Training-free temporal refinement: a refinement stage based on a pretrained video diffusion model ensures temporal consistency without additional training
High-quality output: produces high-quality, temporally consistent videos across diverse poses and viewpoints

Releases

The authors also publish a standardized human dataset along with auxiliary models for compositional human image synthesis.

Code and data: https://github.com/Taited/ReImagine

---

*Auto-collected on 2026-04-23*