UniVOE: A Removal-Driven Framework for Interaction-Aware Universal Video Object Removal and Injection

Forum topic · 小凯 · 2026-10-24

Summary

UniVOE is a unified framework that trains video object removal and injection models jointly by leveraging the inverse relationship between the two tasks. The authors first train an effect-aware remover by transferring video inpainting priors, capable of erasing objects along with their interaction effects such as shadows and motion cues. This remover then converts real object-present videos into object-absent counterparts, generating large-scale paired supervision that injection training would otherwise require but is difficult to obtain from real or synthetic data. The injection model is trained in the reverse direction on these pairs to restore object appearance, dynamics, and scene-dependent effects. To go beyond pixel-level reconstruction, a reward-based post-training stage aligns injection outputs with human preference using a proprietary reward model. Experiments show state-of-the-art results on both removal and injection, with notable strength in challenging scenarios involving interaction effects.

Overview

Field: Computer Vision (CV) Authors: Shuzhou Yang, Lanqing Guo, Lingen Li, Weiqi Li, Zhao Zhong, Liefeng Bo, Jian Zhang Published: 2026-10-22 arXiv: 2610.31641

Key Points

Video object removal and injection are reciprocal editing operations between object-present and object-absent video states, but their training difficulty differs substantially.
Removal can effectively exploit video inpainting priors, then be adapted to erase object-related interaction effects (e.g., shadows, reflections, motion-induced changes).
Injection, by contrast, requires large-scale paired supervision to model object appearance, dynamics, and scene-dependent effects. Perfectly aligned real-world paired data is hard to collect, and synthetic data lacks realism and diversity.

Method

UniVOE is a removal-driven framework that trains both video object removal and injection models jointly:

1. Effect-aware remover: First trained by transferring video inpainting priors, then fine-tuned on object-present/object-absent paired data to erase objects together with their interaction effects. 2. Automatic paired supervision: The trained remover converts real object-present videos into object-absent counterparts, creating large-scale paired training data. 3. Injection model: Trained on these pairs in the reverse direction to recover objects along with their dynamics and interaction effects. 4. Reward-based post-training: Since pixel-level reconstruction alone cannot fully capture perceptual realism, the injection model is further aligned with human preference using a proprietary reward model.

Results

Experiments show that UniVOE achieves state-of-the-art performance on both video object removal and injection tasks, with particularly strong results in challenging scenarios involving interaction effects.