DXcam

Fastest Python Screenshot for Windows

import dxcam
camera = dxcam.create()
camera.grab()

Introduction

DXcam is a Python high-performance screenshot library for Windows using Desktop Duplication API. Capable of 240Hz+ capturing. It was originally built as a part of deep learning pipeline for FPS games to perform better than existed python solutions (python-mss, D3DShot).

Compared to these existed solutions, DXcam provides:

• Way faster screen capturing speed (> 240Hz)
• Capturing of Direct3D exclusive full-screen application without interrupting, even when alt+tab.
• Automatic handling of scaled / stretched resolution.
• Accurate FPS targeting when in capturing mode, makes it suitable for Video output.
• Seamless integration with NumPy, OpenCV, PyTorch, etc.

Contributions are welcome!

Installation

From PyPI:

pip install dxcam

Note: OpenCV is required by DXcam for colorspace conversion. If you don't already have OpenCV, install it easily with command pip install dxcam[cv2].

From source:

pip install --editable .

# for installing OpenCV also
pip install --editable .[cv2]

Usage

In DXCam, each output (monitor) is asscociated to a DXCamera instance. To create a DXCamera instance:

import dxcam
camera = dxcam.create()  # returns a DXCamera instance on primary monitor

Screenshot

For screenshot, simply use .grab:

frame = camera.grab()

The returned frame will be a numpy.ndarray in the shape of (Height, Width, 3[RGB]). This is the default and the only supported format (for now). It is worth noting that .grab will return None if there is no new frame since the last time you called .grab. Usually it means there's nothing new to render since last time (E.g. You are idling).

To view the captured screenshot:

from PIL import Image
Image.fromarray(frame).show()

To screenshot a specific region, use the region parameter: it takes tuple[int, int, int, int] as the left, top, right, bottom coordinates of the bounding box. Similar to PIL.ImageGrab.grab.

left, top = (1920 - 640) // 2, (1080 - 640) // 2
right, bottom = left + 640, top + 640
region = (left, top, right, bottom)
frame = camera.grab(region=region)  # numpy.ndarray of size (640x640x3) -> (HXWXC)

The above code will take a screenshot of the center 640x640 portion of a 1920x1080 monitor.

Screen Capture

To start a screen capture, simply use .start: the capture will be started in a separated thread, default at 60Hz. Use .stop to stop the capture.

camera.start(region=(left, top, right, bottom))  # Optional argument to capture a region
camera.is_capturing  # True
# ... Do Something
camera.stop()
camera.is_capturing  # False

Consume the Screen Capture Data

While the DXCamera instance is in capture mode, you can use .get_latest_frame to get the latest frame in the frame buffer:

camera.start()
for i in range(1000):
    image = camera.get_latest_frame()  # Will block until new frame available
camera.stop()

Notice that .get_latest_frame by default will block until there is a new frame available since the last call to .get_latest_frame. To change this behavior, use video_mode=True.

Advanced Usage and Remarks

Multiple monitors / GPUs

cam1 = dxcam.create(device_idx=0, output_idx=0)
cam2 = dxcam.create(device_idx=0, output_idx=1)
cam3 = dxcam.create(device_idx=1, output_idx=1)
img1 = cam1.grab()
img2 = cam2.grab()
img2 = cam3.grab()

The above code creates three DXCamera instances for: [monitor0, GPU0], [monitor1, GPU0], [monitor1, GPU1], and subsequently takes three full-screen screenshots. (cross GPU untested, but I hope it works.) To get a complete list of devices and outputs:

>>> import dxcam
>>> dxcam.device_info()
'Device[0]:<Device Name:NVIDIA GeForce RTX 3090 Dedicated VRAM:24348Mb VendorId:4318>\n'
>>> dxcam.output_info()
'Device[0] Output[0]: Res:(1920, 1080) Rot:0 Primary:True\nDevice[0] Output[1]: Res:(1920, 1080) Rot:0 Primary:False\n'

Output Format

You can specify the output color mode upon creation of the DXCamera instance:

dxcam.create(output_idx=0, output_color="BGRA")

We currently support "RGB", "RGBA", "BGR", "BGRA", "GRAY", with "GRAY being the gray scale. As for the data format, DXCamera only supports numpy.ndarray in shape of (Height, Width, Channels) right now. We will soon add support for other output formats.

Video Buffer

The captured frames will be insert into a fixed-size ring buffer, and when the buffer is full the newest frame will replace the oldest frame. You can specify the max buffer length (defualt to 64) using the argument max_buffer_len upon creation of the DXCamera instance.

camera = dxcam.create(max_buffer_len=512)

Note: Right now to consume frames during capturing there is only get_latest_frame available which assume the user to process frames in a LIFO pattern. This is a read-only action and won't pop the processed frame from the buffer. we will make changes to support various of consuming pattern soon.

Target FPS

To make DXCamera capture close to the user specified target_fps, we used the undocumented CREATE_WAITABLE_TIMER_HIGH_RESOLUTION flag to create a Windows Waitable Timer Object. This is far more accurate (+/- 1ms) than Python (<3.11) time.sleep (min resolution 16ms). The implementation is done through ctypes creating a perodic timer. Python 3.11 used a similar approach¹.

camera.start(target_fps=120)  # Should not be made greater than 160.

However, due to Windows itself is a preemptive OS² and the overhead of Python calls, the target FPS can not be guarenteed accurate when greater than 160. (See Benchmarks)

Video Mode

The default behavior of .get_latest_frame only put newly rendered frame in the buffer, which suits the usage scenario of a object detection/machine learning pipeline. However, when recording a video that is not ideal since we aim to get the frames at a constant framerate: When the video_mode=True is specified when calling .start method of a DXCamera instance, the frame buffer will be feeded at the target fps, using the last frame if there is no new frame available. For example, the following code output a 5-second, 120Hz screen capture:

target_fps = 120
camera = dxcam.create(output_idx=0, output_color="BGR")
camera.start(target_fps=target_fps, video_mode=True)
writer = cv2.VideoWriter(
    "video.mp4", cv2.VideoWriter_fourcc(*"mp4v"), target_fps, (1920, 1080)
)
for i in range(600):
    writer.write(camera.get_latest_frame())
camera.stop()
writer.release()

You can do interesting stuff with libraries like pyav and pynput: see examples/instant_replay.py for a ghetto implementation of instant replay using hot-keys

Safely Releasing of Resource

Upon calling .release on a DXCamera instance, it will stop any active capturing, free the buffer and release the duplicator and staging resource. Upon calling .stop(), DXCamera will stop the active capture and free the frame buffer. If you want to manually recreate a DXCamera instance on the same output with different parameters, you can also manully delete it:

camera1 = dxcam.create(output_idx=0, output_color="BGR")
camera2 = dxcam.create(output_idx=0)  # Not allowed, camera1 will be returned
camera1 is camera2  # True
del camera1
del camera2
camera2 = dxcam.create(output_idx=0)  # Allowed

Benchmarks

For Max FPS Capability:

start_time, fps = time.perf_counter(), 0
cam = dxcam.create()
start = time.perf_counter()
while fps < 1000:
    frame = cam.grab()
    if frame is not None:  # New frame
        fps += 1
end_time = time.perf_counter() - start_time
print(f"{title}: {fps/end_time}")

When using a similar logistic (only captured new frame counts), DXCam, python-mss, D3DShot benchmarked as follow:

	DXcam	python-mss	D3DShot
Average FPS	238.79 🏁	75.87	118.36
Std Dev	1.25	0.5447	0.3224

The benchmark is across 5 runs, with a light-moderate usage on my PC (5900X + 3090; Chrome ~30tabs, VS Code opened, etc.), I used the Blur Buster UFO test to constantly render 240 fps on my monitor (Zowie 2546K). DXcam captured almost every frame rendered.

For Targeting FPS:

camera = dxcam.create(output_idx=0)
camera.start(target_fps=60)
for i in range(1000):
    image = camera.get_latest_frame()
camera.stop()

(Target)\(mean,std)	DXcam	python-mss	D3DShot
60fps	61.71, 0.26 🏁	N/A	47.11, 1.33
30fps	30.08, 0.02 🏁	N/A	21.24, 0.17

Work Referenced

D3DShot : DXcam borrows the ctypes header directly from the no-longer maintained D3DShot.

OBS Studio : Learned a lot from it.

Footnotes

1. https://github.com/python/cpython/issues/65501 bpo-21302: time.sleep() uses waitable timer on Windows ↩

2. https://en.wikipedia.org/wiki/Preemption_(computing) Preemption (computing) ↩