ECCV 2026
Qinming Zhou1,2,*, Chenxi Sun1,3,*, Deyang Kong1,3, Junhao He1, Xiangheng Tang1,4, Peike Yu1,5, Haotian Wu1, Leilei Cao6, Linfeng Zhang1
1Shanghai Jiao Tong University, 2Tsinghua University, 3University of Electronic Science and Technology of China, 4Xidian University, 5Tongji University, 6Transsion
OSOR is a one-step diffusion framework for efficient, effect-aware, and mask-robust object removal. It removes both the target object and object-associated effects such as shadows, reflections, and residual traces, while running in a single denoising step.
Real-world object removal is challenging due to two key difficulties: the target object's non-local effects, such as shadows and reflections, which are difficult to model, and the fact that user-provided masks are often inaccurate or incomplete. With billions of parameters and tens of denoising steps, diffusion-based models achieve strong removal performance at the expense of substantial computational cost, limiting their use in interactive applications and on edge devices. To address these challenges, we present OSOR (One-Step Object Removal), which simultaneously achieves efficient, effect-aware, and mask-robust object removal. Concretely, OSOR introduces: (1) an occupancy-guided discriminator for precise boundary supervision, enabling stable single-step diffusion training; (2) an alpha head that leverages knowledge from pretrained diffusion models to predict appropriate removal regions with minimal overhead, thereby handling imperfect masks; and (3) a semantic-anchored verification pipeline (SAVP) that filters noisy instruction-based triplets to produce effect-aware supervision at scale. Using SAVP, we curate CORNE, which contains 280K verified removal pairs, and further annotate AnimeEraseBench and TextEraseBench to evaluate performance on more complex removal tasks. Experiments show that OSOR surpasses strong multi-step diffusion baselines in perceptual quality while achieving 4x to 30x faster inference.
- 2026.06: OSOR is available on arXiv.
- 2026.06: We release the open-source code, model weights, CORNE training data, and three evaluation benchmarks.
- Code: SAVP, OSOR-FLUX-Fill, and OSOR-SDXL-Inpainting are released in this repository.
- Paper: arXiv:2606.28094.
- Training data: CORNE on ModelScope.
- Benchmarks: CORNE-Val, AnimeEraseBench, and TextEraseBench on Hugging Face.
- Model weights: QinmingZhou/OSOR on Hugging Face.
| Resource | Link | Description |
|---|---|---|
| Paper | arXiv:2606.28094 | OSOR paper. |
| Model weights | QinmingZhou/OSOR | Released OSOR-FLUX-Fill and OSOR-SDXL-Inpainting checkpoints. |
| Training data | CORNE | 280K verified object-and-effect removal pairs curated with SAVP. |
| Benchmark | CORNE-Val | Paired-background validation benchmark for effect-aware object removal. |
| Benchmark | AnimeEraseBench | Anime-domain benchmark for object-and-effect removal. |
| Benchmark | TextEraseBench | Text removal benchmark for cross-domain evaluation. |
.
├── SAVP/ # Semantic-Anchored Verification Pipeline
├── osor-fluxfill/ # OSOR with the FLUX-Fill backbone
├── osor-sdxlinpainting/ # OSOR with the SDXL-Inpainting backbone
└── assets/
└── first.png
SAVP extracts reliable object-removal supervision from noisy instruction-based editing triplets. It verifies localized semantic changes with GroundingDINO, synthesizes effect-aware masks with perceptual difference analysis and SAM2, and derives object-core masks for incomplete-mask conditioning.
See SAVP/README.md for installation, NHR-Edit preparation, multi-GPU processing, object-core extraction, and reproduction constants.
Both backbone releases follow the same two-phase curriculum:
- Phase I trains one-step removal with hard latent blending and occupancy-guided discriminator supervision.
- Phase II adds alpha prediction and trains with incomplete-mask conditioning for robust effect removal.
Backbone-specific instructions:
- osor-fluxfill: FLUX-Fill implementation.
- osor-sdxlinpainting: SDXL-Inpainting implementation.
Each package contains configs/, scripts/, and src/ and can be used independently.
The released OSOR-FLUX-Fill and OSOR-SDXL-Inpainting code was tested with Python 3.10, PyTorch 2.6.0, and CUDA 12.4. Create a dedicated environment and install the shared dependencies from the repository root:
conda create -n osor python=3.10 -y
conda activate osor
pip install torch==2.6.0 torchvision==0.21.0 --index-url https://download.pytorch.org/whl/cu124
pip install -r requirements.txtFor a different CUDA version, install the matching PyTorch 2.6.0 wheel from PyTorch before running pip install -r requirements.txt. SAVP has a separate dependency stack because it relies on MMDetection and SAM2; follow SAVP/README.md for its tested setup.
Download released checkpoints with:
hf download QinmingZhou/OSOR \
--include "osor-fluxfill/weights/*.pth" \
--local-dir .
hf download QinmingZhou/OSOR \
--include "osor-sdxlinpainting/weights/*.pth" \
--local-dir .The released checkpoint files are:
osor-fluxfill/weights/fluxfill_phase1.pth
osor-fluxfill/weights/fluxfill_phase2.pth
osor-sdxlinpainting/weights/sdxlinpainting_phase1.pth
osor-sdxlinpainting/weights/sdxlinpainting_phase2.pth
If you find this repository, model, or dataset useful for your research, please cite:
@inproceedings{zhou2026osor,
title = {OSOR: One-Step Diffusion Inpainting for Effect-Aware Object Removal},
author = {Zhou, Qinming and Sun, Chenxi and Kong, Deyang and He, Junhao and Tang, Xiangheng and Yu, Peike and Wu, Haotian and Cao, Leilei and Zhang, Linfeng},
booktitle = {European Conference on Computer Vision (ECCV)},
year = {2026},
url = {https://arxiv.org/abs/2606.28094}
}This project builds on open-source diffusion, inpainting, segmentation, and detection toolchains. We thank the authors and maintainers of these projects.