Welcome to visprompt, a repository and basic GUI for experimenting with visual prompting based on SAM and segGPT.

If you would like to learn more about visual prompting, please check out the website accompanying this project.

Current version: 0.1.5

There are two ways in which you can install this repository:

1. standalone
2. add as a dependency in your poetry project

First, clone the project by running:

cd /home/folder/git/
git clone https://github.com/MSchnei/visprompt.git

Then set up a poetry environment by running:

cd /home/folder/git/visprompt/
poetry shell
poetry install

To add visprompt as a dependency to your poetry project, simply run:

poetry add visprompt

To add visprompt as a dependency using pip, simply run:

pip install visprompt

There are two modes in which you can use visprompt:

1. run and visualise segmentations via the GUI
2. run SAM and segGPT segmentation via commmand line

To start the GUI from your terminal, run:

poetry run task gui

Alternatively, to start the GUI from a python shell, run:

from visprompt import run_gui

run_gui()

Once the GUI opens

• drop one or several image(s) for sam segmentation in the top-left window and draw a prompt per image
• drop one or several image(s) for segGPT segmentation in the bottom-left window panel
• click the Submit button

Running the application for the first time might take a while, since we need to download the models from the huggingface hub.

To run a SAM segmentation from your terminal, run the following:

poetry run task inference_sam --prompt-image /path/to/prompt_image.png -p 100 - p 150

To run a segGPT segmentation from your terminal, run

poetry run task inference_seggpt --input-image /path/to/input_image.png --prompt-images /path/to/prompt_image.png --prompt-targets /path/to/prompt_targets.png 

Alternatively, run the below from a python shell:

from PIL import Image
from visprompt import SAMInference, SegGPTInference

# Set prompt_image and input_points for SAM segmentation
prompt_image = Image.open("/path/to/prompt_image.png").convert("RGB")
input_points = [[[100, 150]]]

# Run SAM segmentation
inference_instance = SAMInference()
mask = inference_instance.run_inference(
    prompt_image=prompt_image,
    input_points=input_points,
)

# Set input_image, prompt_images and prompt_targets for SegGPT segmentation
input_image = Image.open("/path/to/input_image.png").convert("RGB")
prompt_images = [Image.open("/path/to/prompt_image.png").convert("RGB")]
prompt_targets = [Image.open("/path/to/prompt_target.png").convert("RGB")]

# Run SegGPT segmentation
inference_instance = SegGPTInference(num_labels=1)
mask = inference_instance.run_inference(
    input_image=input_image,
    prompt_images=prompt_images,
    prompt_targets=prompt_targets,
)

Contributions are welcome! Before submitting a PR, please run:

make style

This will run black, isort and flake8 on the code.

Unit tests can be executed via

make test