---
title: "OmniPaint: Mastering Object-Oriented Editing via Disentangled Insertion-Removal Inpainting"
canonical_url: "https://www.modelscope.cn/papers/125491"
md_url: "https://www.modelscope.cn/papers/125491.md"
arxiv_id: 2503.08677
published: 2025-03-11
last_updated: 2025-03-11
authors:
  - "Yongsheng Yu"
  - "Ziyun Zeng"
  - "Haitian Zheng"
  - "Jiebo Luo"
model_name: OmniPaint
model_developer: "罗切斯特大学，Adobe研究院"
domain:
  - "计算机视觉"
  - "深度学习"
type:
  - "计算机视觉"
  - "深度学习"
  - "Computer Vision and Pattern Recognition (cs.CV)"
arxiv_url: "https://arxiv.org/abs/2503.08677"
pdf_url: "https://arxiv.org/pdf/2503.08677.pdf"
code_link: "https://github.com/yeates/OmniPaint-Page/"
---

# OmniPaint: Mastering Object-Oriented Editing via Disentangled Insertion-Removal Inpainting

> Diffusion-based generative models have revolutionized object-oriented image editing, yet their deployment in realistic object removal and insertion remains hampered by challenges such as the intricate interplay of physical effects and insufficient paired…

「OmniPaint: Mastering Object-Oriented Editing via Disentangled Insertion-Removal Inpainting」是 ModelScope 魔搭社区收录的论文，arXiv 2503.08677，作者为 Yongsheng Yu, Ziyun Zeng, Haitian Zheng et al.，发表于 2025-03-11，属于 计算机视觉、深度学习 领域。

- **ArXiv**: 2503.08677
- **Published**: 2025-03-11
- **Authors**: Yongsheng Yu, Ziyun Zeng, Haitian Zheng, Jiebo Luo
- **Model**: OmniPaint
- **Developer**: 罗切斯特大学，Adobe研究院
- **Domain**: 计算机视觉, 深度学习
- **ArXiv URL**: https://arxiv.org/abs/2503.08677
- **PDF**: https://arxiv.org/pdf/2503.08677.pdf
- **Code**: https://github.com/yeates/OmniPaint-Page/

Source: https://www.modelscope.cn/papers/125491

---

> OmniPaint：通过解耦插入与移除修复实现对象级图像编辑的新突破

## 摘要

本文提出了一种名为OmniPaint的统一框架，旨在通过解耦插入-移除修复（disentangled insertion-removal inpainting）来实现面向对象编辑的任务。传统方法通常将对象移除和插入视为独立任务，而OmniPaint将其重新概念化为相互依赖的过程。该框架基于预训练扩散先验（如FLUX），并通过渐进式训练管道实现优化，包括初始配对样本优化和后续的大规模非配对精炼（CycleFlow）。OmniPaint能够在精确消除前景的同时保留场景几何和内在属性，并在对象插入时实现无缝融合。此外，作者还提出了一个新的无参考评估指标CFD（Context-Aware Feature Derivation），用于检测上下文一致性与对象幻觉，从而为高保真图像编辑提供了新基准。实验表明，OmniPaint在处理复杂物理效果（如阴影和反射）方面表现出色，同时显著减少了对大规模配对数据集的依赖。

## Abstract

Diffusion-based generative models have revolutionized object-oriented image editing, yet their deployment in realistic object removal and insertion remains hampered by challenges such as the intricate interplay of physical effects and insufficient paired training data. In this work, we introduce OmniPaint, a unified framework that re-conceptualizes object removal and insertion as interdependent processes rather than isolated tasks. Leveraging a pre-trained diffusion prior along with a progressive training pipeline comprising initial paired sample optimization and subsequent large-scale unpaired refinement via CycleFlow, OmniPaint achieves precise foreground elimination and seamless object insertion while faithfully preserving scene geometry and intrinsic properties. Furthermore, our novel CFD metric offers a robust, reference-free evaluation of context consistency and object hallucination, establishing a new benchmark for high-fidelity image editing. Project page: this https URL