gui-agent-screenshot-tools 1.0.0

Screenshot, bounding-box, and coordinate transformation utilities for GUI agents and other GUI related tasks.

gui-agent-screenshot-tools

Coordinate, bounding-box, and screenshot resize utilities for GUI agents.

This library helps you map detections between:

• original screenshot space
• resized screenshot space (letterbox or stretch)
• physical screen space (with optional window offsets)

Installation

pip install gui-agent-screenshot-tools

Quick Start

from gui_agent_screenshot_tools import Coordinate, Space

source = Space(width=1920, height=1080)
target = Space(width=1280, height=720)

coord = Coordinate(x=960, y=540, space=source)
mapped = coord.to_space(target)
print(mapped.x, mapped.y)  # 640 360

Directional APIs

Use directional methods when you have resize metadata:

• to_source_space(...): resized -> original
• to_target_space(...): original -> resized

Compatibility aliases are still available:

• Coordinate.to_space(..., resize_metadata=...)
• ResizeMetadata.transform_coordinate(...)
• ResizeMetadata.forward_transform_coordinate(...)

Workflow 1: Detection In Resized Image -> Original Coordinates

from PIL import Image
from gui_agent_screenshot_tools import Coordinate, ResizeMode, Screenshot, Space

img = Image.new("RGB", (1920, 1080), color=(255, 0, 0))
original = Screenshot.from_image(img)

resized = original.resize(Space(width=1024, height=1024), ResizeMode.LETTERBOX)
model_point = Coordinate(x=512, y=512, space=resized.space)

# New helper API
point_in_original = resized.map_coord_to_original(model_point)
print(point_in_original.x, point_in_original.y)  # 960 540

Workflow 2: Window-Local BBox -> Screen Coordinates

from PIL import Image
from gui_agent_screenshot_tools import BBox, ResizeMode, Screenshot, Space

screen = Space(width=2560, height=1440)
window = BBox(x=50, y=100, width=1920, height=1080, space=screen)

# Screenshot is of the window content only
img = Image.new("RGB", (1920, 1080), color=(80, 120, 180))
original = Screenshot.from_image(img)
resized = original.resize(Space(width=1024, height=1024), ResizeMode.LETTERBOX)

# Model output in resized screenshot space
model_bbox = BBox(x=100, y=200, width=300, height=150, space=resized.space)

# One call: resized -> original -> screen (via window offset)
screen_bbox = resized.map_bbox_to_original(model_bbox, offset=window)
print(screen_bbox.x, screen_bbox.y, screen_bbox.width, screen_bbox.height)
# 238 100 561 236

Common Mistakes

• Using a coordinate with the wrong space for the metadata direction. Example: to_source_space expects coordinates in metadata.target_space.
• Calling screenshot remap helpers before resize(...). map_coord_* and map_bbox_* require resize_metadata.
• Passing an offset bbox whose offset.as_space does not match the post-transform local space.
• Passing a raw string mode to resize. Use ResizeMode.LETTERBOX or ResizeMode.STRETCH.

Development

uv run --extra dev python -m pytest -q

License

MIT