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

ICCV 2025

       

Features

• 🧹 Object Removal - Remove foreground objects and their effects using only object masks
• 🖼️ Object Insertion - Seamless generative insertion of objects into existing scenes
• 📊 Novel CFD Metric - Reference-free evaluation of object removal quality

Setup

Install dependencies and download OmniPaint weights:

bash scripts/setup.sh

Run this from the repository root.

Code Usage

Object Removal

• Single image:

python scripts/omnipaint_remove.py \
  --input ./demo_assets/removal_samples/images/5.jpg \
  --mask ./demo_assets/removal_samples/masks/5.png \
  --output_dir ./outputs \
  --seed 42 \
  --steps 28 \
  --device cuda:0

• Directory:

python scripts/omnipaint_remove.py \
  --input ./demo_assets/removal_samples/images \
  --mask ./demo_assets/removal_samples/masks \
  --output_dir ./outputs \
  --seed 42 \
  --steps 28 \
  --device cuda:0

Object Insertion

• Single image:

python scripts/omnipaint_insert.py \
  --background ./demo_assets/insertion_samples/backgrounds/background-2.png \
  --mask ./demo_assets/insertion_samples/masks/mask-2.png \
  --subject ./demo_assets/insertion_samples/subjects/subject-2.png \
  --output_dir ./outputs \
  --seed 42 \
  --steps 28 \
  --device cuda:0 \
  --carvekit_device cuda:0

• Directory:

python scripts/omnipaint_insert.py \
  --background ./demo_assets/insertion_samples/backgrounds \
  --mask ./demo_assets/insertion_samples/masks \
  --subject ./demo_assets/insertion_samples/subjects \
  --output_dir ./outputs \
  --seed 42 \
  --steps 28 \
  --device cuda:0 \
  --carvekit_device cuda:0

Demo App

The demo app supports both manual mask drawing and automatic mask generation using segmentation model.

• Gradio environment setup:

  bash scripts/app_setup.sh
  

  Run this if you plan to launch app.py. The installation may take around 15 minutes due to SAM2 setup and weight downloads.

• Run the app:

  python app.py
  

Notes

• Directory mode expects the following structure by default:
  • Removal: demo_assets/removal_samples/images/*.{jpg,png,...} with matching masks in demo_assets/removal_samples/masks/*.{jpg,png,...} (same basenames).
  • Insertion: demo_assets/insertion_samples/backgrounds/background-XX.png, masks in demo_assets/insertion_samples/masks/mask-XX.png, subjects in demo_assets/insertion_samples/subjects/subject-XX.png. Direct basename alignment also works.
• Mask quality strongly affects insertion performance. Prefer a single connected-component mask; avoid multiple disconnected masks.

Evaluation - CFD Score

See cfd_score for setup and usage.

Citation

If you find OmniPaint useful, please cite our paper:

@inproceedings{yu2025omnipaint,
  title={Omnipaint: Mastering object-oriented editing via disentangled insertion-removal inpainting},
  author={Yu, Yongsheng and Zeng, Ziyun and Zheng, Haitian and Luo, Jiebo},
  booktitle={ICCV},
  pages={17324--17334},
  year={2025}
}