Skip to content

Latest commit

 

History

History
433 lines (372 loc) · 20.4 KB

README.md

File metadata and controls

433 lines (372 loc) · 20.4 KB

Inpaint Anything: Segment Anything Meets Image Inpainting

Inpaint Anything can inpaint anything in images, videos and 3D scenes!

  • Authors: Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng and Zhibo Chen.
  • Institutes: University of Science and Technology of China; Eastern Institute for Advanced Study.
  • [Paper] [Website] [Hugging Face Homepage]

TL; DR: Users can select any object in an image by clicking on it. With powerful vision models, e.g., SAM, LaMa and Stable Diffusion (SD), Inpaint Anything is able to remove the object smoothly (i.e., Remove Anything). Further, prompted by user input text, Inpaint Anything can fill the object with any desired content (i.e., Fill Anything) or replace the background of it arbitrarily (i.e., Replace Anything).

📜 News

[2023/4/30] Remove Anything Video available! You can remove any object from a video!
[2023/4/24] Local web UI supported! You can run the demo website locally!
[2023/4/22] Website available! You can experience Inpaint Anything through the interface!
[2023/4/22] Remove Anything 3D available! You can remove any 3D object from a 3D scene!
[2023/4/13] Technical report on arXiv available!

🌟 Features

💡 Highlights

  • Any aspect ratio supported
  • 2K resolution supported
  • Technical report on arXiv available (🔥NEW)
  • Website available (🔥NEW)
  • Local web UI available (🔥NEW)
  • Multiple modalities (i.e., image, video and 3D scene) supported (🔥NEW)

📌 Remove Anything

image

Click on an object in the image, and Inpainting Anything will remove it instantly!

Installation

Requires python>=3.8

python -m pip install torch torchvision torchaudio
python -m pip install -e segment_anything
python -m pip install -r lama/requirements.txt 

In Windows, we recommend you to first install miniconda and open Anaconda Powershell Prompt (miniconda3) as administrator. Then pip install ./lama_requirements_windows.txt instead of ./lama/requirements.txt.

Usage

Download the model checkpoints provided in Segment Anything and LaMa (e.g., sam_vit_h_4b8939.pth and big-lama), and put them into ./pretrained_models. For simplicity, you can also go here, directly download pretrained_models, put the directory into ./ and get ./pretrained_models.

Specify an image and a point, and Remove Anything will remove the object at the point.

python remove_anything.py \
    --input_img ./example/remove-anything/dog.jpg \
    --coords_type key_in \
    --point_coords 200 450 \
    --point_labels 1 \
    --dilate_kernel_size 15 \
    --output_dir ./results \
    --sam_model_type "vit_h" \
    --sam_ckpt ./pretrained_models/sam_vit_h_4b8939.pth \
    --lama_config ./lama/configs/prediction/default.yaml \
    --lama_ckpt ./pretrained_models/big-lama

You can change --coords_type key_in to --coords_type click if your machine has a display device. If click is set, after running the above command, the image will be displayed. (1) Use left-click to record the coordinates of the click. It supports modifying points, and only last point coordinates are recorded. (2) Use right-click to finish the selection.

Demo

📌 Fill Anything

Text prompt: "a teddy bear on a bench"

image

Click on an object, type in what you want to fill, and Inpaint Anything will fill it!

  • Click on an object;
  • SAM segments the object out;
  • Input a text prompt;
  • Text-prompt-guided inpainting models (e.g., Stable Diffusion) fill the "hole" according to the text.

Installation

Requires python>=3.8

python -m pip install torch torchvision torchaudio
python -m pip install -e segment_anything
python -m pip install diffusers transformers accelerate scipy safetensors

Usage

Download the model checkpoints provided in Segment Anything (e.g., sam_vit_h_4b8939.pth) and put them into ./pretrained_models. For simplicity, you can also go here, directly download pretrained_models, put the directory into ./ and get ./pretrained_models.

Specify an image, a point and text prompt, and run:

python fill_anything.py \
    --input_img ./example/fill-anything/sample1.png \
    --coords_type key_in \
    --point_coords 750 500 \
    --point_labels 1 \
    --text_prompt "a teddy bear on a bench" \
    --dilate_kernel_size 50 \
    --output_dir ./results \
    --sam_model_type "vit_h" \
    --sam_ckpt ./pretrained_models/sam_vit_h_4b8939.pth

Demo

Text prompt: "a camera lens in the hand"
Text prompt: "a Picasso painting on the wall"
Text prompt: "an aircraft carrier on the sea"
Text prompt: "a sports car on a road"

📌 Replace Anything

Text prompt: "a man in office"

image

Click on an object, type in what background you want to replace, and Inpaint Anything will replace it!

  • Click on an object;
  • SAM segments the object out;
  • Input a text prompt;
  • Text-prompt-guided inpainting models (e.g., Stable Diffusion) replace the background according to the text.

Installation

Requires python>=3.8

python -m pip install torch torchvision torchaudio
python -m pip install -e segment_anything
python -m pip install diffusers transformers accelerate scipy safetensors

Usage

Download the model checkpoints provided in Segment Anything (e.g. sam_vit_h_4b8939.pth) and put them into ./pretrained_models. For simplicity, you can also go here, directly download pretrained_models, put the directory into ./ and get ./pretrained_models.

Specify an image, a point and text prompt, and run:

python replace_anything.py \
    --input_img ./example/replace-anything/dog.png \
    --coords_type key_in \
    --point_coords 750 500 \
    --point_labels 1 \
    --text_prompt "sit on the swing" \
    --output_dir ./results \
    --sam_model_type "vit_h" \
    --sam_ckpt ./pretrained_models/sam_vit_h_4b8939.pth

Demo

Text prompt: "sit on the swing"
Text prompt: "a bus, on the center of a country road, summer"
Text prompt: "breakfast"
Text prompt: "crossroad in the city"

📌 Remove Anything 3D

Remove Anything 3D can remove any object from a 3D scene! We release some results below. (Code and implementation details will be released soon.)

📌 Remove Anything Video

With a single click on an object in the first video frame, Remove Anything Video can remove the object from the whole video!

  • Click on an object in the first frame of a video;
  • SAM segments the object out (with three possible masks);
  • Select one mask;
  • A tracking model such as OSTrack is ultilized to track the object in the video;
  • SAM segments the object out in each frame according to tracking results;
  • A video inpainting model such as STTN is ultilized to inpaint the object in each frame.

Installation

Requires python>=3.8

python -m pip install torch torchvision torchaudio
python -m pip install -e segment_anything
python -m pip install -r lama/requirements.txt
python -m pip install jpeg4py lmdb

Usage

Download the model checkpoints provided in Segment Anything and STTN (e.g., sam_vit_h_4b8939.pth and sttn.pth), and put them into ./pretrained_models. Further, download OSTrack pretrained model from here (e.g., vitb_384_mae_ce_32x4_ep300.pth) and put it into ./pytracking/pretrain. For simplicity, you can also go here, directly download pretrained_models, put the directory into ./ and get ./pretrained_models. Additionally, download pretrain, put the directory into ./pytracking and get ./pytracking/pretrain.

Specify a video, a point, video FPS and mask index (indicating using which mask result of the first frame), and Remove Anything Video will remove the object from the whole video.

python remove_anything_video.py \
    --input_video ./example/video/paragliding/original_video.mp4 \
    --coords_type key_in \
    --point_coords 652 162 \
    --point_labels 1 \
    --dilate_kernel_size 15 \
    --output_dir ./results \
    --sam_model_type "vit_h" \
    --sam_ckpt ./pretrained_models/sam_vit_h_4b8939.pth \
    --lama_config lama/configs/prediction/default.yaml \
    --lama_ckpt ./pretrained_models/big-lama \
    --tracker_ckpt vitb_384_mae_ce_32x4_ep300 \
    --vi_ckpt ./pretrained_models/sttn.pth \
    --mask_idx 2 \
    --fps 25

The --mask_idx is usually set to 2, which typically is the most confident mask result of the first frame. If the object is not segmented out well, you can try other masks (0 or 1).

Demo

Acknowledgments

Other Interesting Repositories

Citation

If you find this work useful for your research, please cite us:

@article{yu2023inpaint,
  title={Inpaint Anything: Segment Anything Meets Image Inpainting},
  author={Yu, Tao and Feng, Runseng and Feng, Ruoyu and Liu, Jinming and Jin, Xin and Zeng, Wenjun and Chen, Zhibo},
  journal={arXiv preprint arXiv:2304.06790},
  year={2023}
}

Star History Chart