Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen and Fang Wen.

Language-guided image editing has achieved great success recently. In this paper, for the first time, we investigate exemplar-guided image editing for more precise control. We achieve this goal by leveraging self-supervised training to disentangle and re-organize the source image and the exemplar. However, the naive approach will cause obvious fusing artifacts. We carefully analyze it and propose an information bottleneck and strong augmentations to avoid the trivial solution of directly copying and pasting the exemplar image. Meanwhile, to ensure the controllability of the editing process, we design an arbitrary shape mask for the exemplar image and leverage the classifier-free guidance to increase the similarity to the exemplar image. The whole framework involves a single forward of the diffusion model without any iterative optimization. We demonstrate that our method achieves an impressive performance and enables controllable editing on in-the-wild images with high fidelity.

• 2023-11-28 The recent work Asymmetric VQGAN improves the preservation of details in non-masked regions. For comprehensive details, please refer to the associated paper, github.
• 2023-05-13 Release code for quantitative results.
• 2023-03-03 Release test benchmark.
• 2023-02-23 Non-official 3rd party apps support by ModelScope (the largest Model Community in Chinese).
• 2022-12-07 Release a Gradio demo on Hugging Face Spaces.
• 2022-11-29 Upload code.

A suitable conda environment named Paint-by-Example can be created and activated with:

conda env create -f environment.yaml
conda activate Paint-by-Example

We provide the checkpoint (Google Drive | Hugging Face) that is trained on Open-Images for 40 epochs. By default, we assume that the pretrained model is downloaded and saved to the directory checkpoints.

To sample from our model, you can use scripts/inference.py. For example,

python scripts/inference.py \
--plms --outdir results \
--config configs/v1.yaml \
--ckpt checkpoints/model.ckpt \
--image_path examples/image/example_1.png \
--mask_path examples/mask/example_1.png \
--reference_path examples/reference/example_1.jpg \
--seed 321 \
--scale 5

or simply run:

sh test.sh

Visualization of inputs and output:

• Download separate packed files of Open-Images dataset from CVDF's site and unzip them to the directory dataset/open-images/images.
• Download bbox annotations of Open-Images dataset from Open-Images official site and save them to the directory dataset/open-images/annotations.
• Generate bbox annotations of each image in txt format.

  python scripts/read_bbox.py
  

The data structure is like this:

dataset
├── open-images
│  ├── annotations
│  │  ├── class-descriptions-boxable.csv
│  │  ├── oidv6-train-annotations-bbox.csv
│  │  ├── test-annotations-bbox.csv
│  │  ├── validation-annotations-bbox.csv
│  ├── images
│  │  ├── train_0
│  │  │  ├── xxx.jpg
│  │  │  ├── ...
│  │  ├── train_1
│  │  ├── ...
│  │  ├── validation
│  │  ├── test
│  ├── bbox
│  │  ├── train_0
│  │  │  ├── xxx.txt
│  │  │  ├── ...
│  │  ├── train_1
│  │  ├── ...
│  │  ├── validation
│  │  ├── test

We utilize the pretrained Stable Diffusion v1-4 as initialization, please download the pretrained models from Hugging Face and save the model to directory pretrained_models. Then run the following script to add zero-initialized weights for 5 additional input channels of the UNet (4 for the encoded masked-image and 1 for the mask itself).

python scripts/modify_checkpoints.py

To train a new model on Open-Images, you can use main.py. For example,

python -u main.py \
--logdir models/Paint-by-Example \
--pretrained_model pretrained_models/sd-v1-4-modified-9channel.ckpt \
--base configs/v1.yaml \
--scale_lr False

or simply run:

sh train.sh

We build a test benchmark for quantitative analysis. Specifically, we manually select 3500 source images from MSCOCO validation set, each image contains only one bounding box. Then we manually retrieve a reference image patch from MSCOCO training set. The reference image usually shares a similar semantic with mask region to ensure the combination is reasonable. We named it as COCO Exemplar-based image Editing benchmark, abbreviated as COCOEE. This test benchmark can be downloaded from Google Drive.

By default, we assume that the COCOEE is downloaded and saved to the directory test_bench. To generate the results of test bench, you can use scripts/inference_test_bench.py. For example,

python scripts/inference_test_bench.py \
--plms \
--outdir results/test_bench \
--config configs/v1.yaml \
--ckpt checkpoints/model.ckpt \
--scale 5

or simply run:

bash inference_test_bench.sh

By default, we assume that the test set of COCO2017 is downloaded and saved to the directory dataset. The data structure is like this:

dataset
├── coco
│  ├── test2017
│  │  ├── xxx.jpg
│  │  ├── xxx.jpg
│  │  ├── ...
│  │  ├── xxx.jpg

Then convert the images into square images with 512 solution.

python scripts/create_square_gt_for_fid.py

To calculate FID score, simply run:

python eval_tool/fid/fid_score.py --device cuda \
test_bench/test_set_GT \
results/test_bench/results

Please download the model weights for QS score from Google Drive and save the model to directory eval_tool/gmm. To calculate QS score, simply run:

python eval_tool/gmm/gmm_score_coco.py results/test_bench/results \
--gmm_path eval_tool/gmm/coco2017_gmm_k20 \
--gpu 1

To calculate CLIP score, simply run:

python eval_tool/clip_score/region_clip_score.py \
--result_dir results/test_bench/results

@article{yang2022paint,
  title={Paint by Example: Exemplar-based Image Editing with Diffusion Models},
  author={Binxin Yang and Shuyang Gu and Bo Zhang and Ting Zhang and Xuejin Chen and Xiaoyan Sun and Dong Chen and Fang Wen},
  journal={arXiv preprint arXiv:2211.13227},
  year={2022}
}

This code borrows heavily from Stable Diffusion. We also thank the contributors of OpenAI's ADM codebase and https://github.com/lucidrains/denoising-diffusion-pytorch.

Please open a GitHub issue for any help. If you have any questions regarding the technical details, feel free to contact us.

The codes and the pretrained model in this repository are under the CreativeML OpenRAIL M license as specified by the LICENSE file.

The test benchmark, COCOEE, belongs to the COCO Consortium and are licensed under a Creative Commons Attribution 4.0 License.