Precise Object and Effect Removal with Adaptive Target-Aware Attention

Jixin Zhao^*  Zhouxia Wang  Peiqing Yang  Shangchen Zhou^*,†

S-Lab, Nanyang Technological University 

CVPR 2026

ObjectClear is an object removal model that can jointly eliminate the target object and its associated effects leveraging Adaptive Target-Aware Attention, while preserving background consistency.

For more visual results, go checkout our project page

---

• [2026.08] 🔥 Training code is now released! See the Training section below.
• [2026.02] 🔥 OBER Dataset is Now Released! Our training dataset is now publicly available on Hugging Face 🤗.
• [2025.09] We have released our benchmark datasets for evaluation, along with our results to facilitate comparison.
• [2025.07] Release the inference code and Gradio demo.
• [2025.05] This repo is created.

• Release the training code
• Release our training datasets
• Release our benchmark datasets
• Release the inference code and Gradio demo

OBER (OBject-Effect Removal) is a hybrid dataset designed to support research in object removal with effects, combining both camera-captured and simulated data.

🔥 We have released the full dataset OBERDataset_ObjectClear on Hugging Face. We hope it can serve as a strong training resource and benchmark for future object removal research.

🚩 Note that the OBER dataset are made available solely for non-commercial research use. Any use, reproduction, or redistribution must strictly comply with the terms of NTU S-Lab License 1.0.

1. Clone Repo

   git clone https://github.com/zjx0101/ObjectClear.git
   cd ObjectClear

2. Create Conda Environment and Install Dependencies

   # create new conda env
   conda create -n objectclear python=3.10 -y
   conda activate objectclear
   
   # install python dependencies
   pip3 install -r requirements.txt
   # [optional] install python dependencies for gradio demo
   pip3 install -r hugging_face/requirements.txt

We provide some examples in the inputs folder. For each run, we take an image and its segmenatation mask as input. The segmentation mask can be obtained from interactive segmentation models such as SAM2 demo. For example, the directory structure can be arranged as follows:

inputs
   ├─ imgs
   │   ├─ test-sample1.jpg      # .jpg, .png, .jpeg supported
   │   ├─ test-sample2.jpg
   └─ masks
       ├─ test-sample1.png
       ├─ test-sample2.png

Run the following command to try it out:

## Single image inference
python inference_objectclear.py -i inputs/imgs/test-sample1.jpg -m inputs/masks/test-sample1.png --guidance_scale 2.5 --use_fp16

## Batch inference on image folder
python inference_objectclear.py -i inputs/imgs -m inputs/masks --guidance_scale 2.5 --use_fp16

Note: --guidance_scale controls the trade-off: higher values lead to stronger removal, while lower values better preserve background details.
The default setting is --guidance_scale 2.5. For all benchmark results reported in our paper, we used --guidance_scale 1.0.

ObjectClear is built on SDXL-Inpainting and uses a CLIP image encoder to encode the target object. The SDXL base model is downloaded automatically on the first run. Download the CLIP image encoder into ./ckpts:

# download the CLIP image encoder used as the object encoder
huggingface-cli download openai/clip-vit-large-patch14 --local-dir ./ckpts/clip-vit-large-patch14

The training reads the OBER parquet shards directly (no unpacking needed). Access to the dataset is gated — first request access on the dataset page, then log in and download:

# log in with your Hugging Face token (needed for the gated dataset)
huggingface-cli login

# download the OBER dataset (~27 GB) into ./data/OBER
huggingface-cli download sczhou/OBERDataset_ObjectClear --repo-type dataset --local-dir ./data/OBER

After downloading, the parquet shards live in ./data/OBER/data:

data/OBER/data
   ├─ train-00000-of-00053.parquet   # 37,994 cropped training pairs
   ├─ ...
   ├─ train-00052-of-00053.parquet
   └─ test-00000-of-00001.parquet    # test split (used for validation)

Each sample provides input, gt, object_mask, and object_effect_mask.

We provide a ready-to-run script train.sh for multi-GPU training with 🤗 accelerate. Edit the paths / hyper-parameters at the top of the script, then run:

bash train.sh

Or launch directly with accelerate (8 GPUs example):

accelerate launch --multi_gpu --num_processes 8 --mixed_precision fp16 \
    train_objectclear.py \
    --pretrained_model_name_or_path "diffusers/stable-diffusion-xl-1.0-inpainting-0.1" \
    --image_encoder_name_or_path "./ckpts/clip-vit-large-patch14" \
    --output_dir "./runs/train_objectclear" \
    --image_dir1 "./data/OBER/data" \
    --resolution 512 \
    --train_batch_size 4 \
    --learning_rate 1e-05 \
    --learning_rate_attn 1e-05 \
    --lr_scheduler cosine \
    --max_train_steps 100000 \
    --checkpointing_steps 5000 \
    --checkpoints_total_limit 5 \
    --color_augmentation \
    --flip_augmentation \
    --random_mask_dilation \
    --random_mask_erosion \
    --object_localization \
    --object_localization_weight 0.01 \
    --background_loss_weight 1 \
    --real_only \
    --seed 42

To monitor training, enable validation on the OBER test split. Validation runs the full ObjectClearPipeline (identical to inference) and reports PSNR on samples with ground truth:

    --validation_parquet "./data/OBER/data/test-00000-of-00001.parquet" \
    --validation_subset "OBER-Test" \
    --validation_num_samples 8 \
    --validate_by_iter \
    --validation_iterations 2000

Note: --validation_subset can be OBER-Test or RORD-Val-343 (both have ground truth, so PSNR is computed) or OBER-Wild (no ground truth). Validation images, attention maps, and metrics are written to <output_dir>/validation_results/.

Our ReMOVE+ metric addresses the limitations of the original ReMOVE by assessing consistency between the output's object-effect region and the input's background (outside the object-effect mask), making it more suitable for object-effect removal evaluation.

Please refer to the detailed instructions in the evaluation/README.md file for installation, setup, and running the ReMOVE+ evaluation pipeline.

To get rid of the preparation for segmentation mask, we prepare a gradio demo on hugging face and could also launch locally. Just drop your image, assign the target masks with a few clicks, and get the object removal results!

cd hugging_face

# install python dependencies
pip3 install -r requirements.txt

# launch the demo
python app.py

Non-Commercial Use Only Declaration

The ObjectClear is made available for use, reproduction, and distribution strictly for non-commercial purposes. The code, models, and datasets are licensed under NTU S-Lab License 1.0. Redistribution and use should follow this license.

If you find our repo useful for your research, please consider citing our paper:

@InProceedings{zhao2026objectclear,
    title   = {Precise Object and Effect Removal with Adaptive Target-Aware Attention},
    author  = {Zhao, Jixin and Wang, Zhouxia and Yang, Peiqing and Zhou, Shangchen},
    booktitle = {CVPR},
    year    = {2026},
    }

If you have any questions, please feel free to reach us at jixinzhao0101@gmail.com and shangchenzhou@gmail.com.