rustcam 0.0.6

Rust-backed Windows DXGI Desktop Duplication API screen capture for Python.

rustcam

Fast DXGI Desktop Duplication screen capture for Windows, in Rust.

I made this because every "fast" screen capture package on PyPI runs its hot loop in Python. bettercam is a fork of dxcam, dxcam calls AcquireNextFrame through comtypes on every frame under the GIL, and the GDI-based ones (mss, PIL.ImageGrab) aren't even using DDA. They all top out around 130-140 fps on a 180 Hz monitor for the same reason: per-frame Python overhead misses compositor ticks. rustcam runs the whole AcquireNextFrame -> CopyResource -> Map -> memcpy cycle in native Rust with the GIL released, so it actually rides the refresh rate.

import rustcam

cap = rustcam.Capturer(output=0, cursor=True)
frame = cap.grab()        # numpy ndarray (H, W, 4) BGRA, or None on timeout

Prebuilt Windows wheels for Python 3.9 through 3.13 (a single abi3 wheel that covers them all). pip install rustcam never compiles anything.

Install

pip install rustcam

Windows only. DDA is the IDXGIOutputDuplication interface, which is Win8+. There is no Linux or macOS equivalent. If you need cross-platform capture, look at mss (slower, GDI-based).

Performance

benches/flip_demo_bench.py runs each library against two stimuli, on a 1920x1080 / 180 Hz monitor backed by a GTX 1660 Ti:

• flip_demo is a native Rust D3D11 app from the zentape project. It presents a flip-model swapchain with a unique full-screen colour every refresh, so the source emits ~180 unique fps and any honest capturer should be able to read at panel rate. This is the controlled benchmark.
• mover.py is a borderless PyQt window that orbits across the screen continuously. This is the realistic benchmark, what users actually capture, dragging a window around or recording a moving UI.

The metric is valid fps — how many non-None frames the lib returns per second, with %changed annotating how often consecutive returned frames actually differ. Earlier versions of this README compared rustcam's normal API against bettercam's .grab() mode, which is a fast-but-unrealistic tight-loop pattern nobody actually codes against. The numbers below use bettercam.start()/.get_latest_frame() and dxcam.start()/.get_latest_frame() — the only modes either library is actually used in.

The harness for these numbers is benches/controlled_bench.py. It pops a status window on the top monitor, gives you 3 seconds to finish whatever you're doing, minimizes all your visible windows (it remembers each HWND so it can restore them after, no Win+D guesswork), runs each capturer in a subprocess so DDA state doesn't leak between trials, takes the median of 3 runs per cell, and restores your windows at the end. Run it yourself, the numbers should reproduce within a few percent.

capturer	flip_demo (valid fps)	mover.py (valid fps)	mover.py (% changed)
rustcam grab(cursor=False)	~176	~175	100 %
rustcam grab(cursor=True)	~175	~175	100 %
rustcam start/get_latest_frame	~174	~173	100 %
bettercam .start()/.get_latest_frame()	~153	~148	100 %
dxcam .start()/.get_latest_frame()	~159	~157	100 %
mss	~56	~54	100 %

(Numbers are from a clean controlled run; medians of 3 trials per cell; tight error bars in single-digit fps.)

On the controlled flip-model source (flip_demo is a tiny native Rust D3D11 app that presents a unique full-screen colour every refresh), rustcam delivers 174-176 valid fps. The ~4 fps deficit vs the 180 Hz panel rate is the cost of the xxhash full-frame fingerprint we run inside the bench loop to count uniques; same cost on every capturer so the comparison stays fair. bettercam's .start() mode tops out around 153, dxcam around 159 — their bg threads are Python loops, ours is native Rust with the GIL released.

On realistic moving content (mover.py is a borderless PyQt window that orbits across the screen), rustcam stays at 173-175 fps, basically matching the controlled source. bettercam drops slightly to ~148, dxcam to ~157. The headline isn't a 5x gap anymore now that we're measuring honestly, but rustcam still wins by 15-25 fps and has tighter error bars (sub-1 fps variance vs bettercam's 9 fps and dxcam's 5 fps). And cursor=True no longer collapses on real content (it was 47 fps in v0.0.5; the fix is in the v0.0.6 section below).

v0.0.6 cursor=True fix

Up through v0.0.5, cursor=True had a hidden bottleneck on realistic moving content. draw_cursor used IDXGISurface1::GetDC on a MISC_GDI_COMPATIBLE texture, which is a CPU↔GPU sync point. When DWM was awake compositing other windows (which is most actual desktop work), the sync queued behind DWM's per-vsync work and capped cursor=True throughput at 47 fps, vs ~150 for cursor=False. flip_demo hid the issue because it triggers Independent Flip, which puts DWM to sleep entirely.

v0.0.6 drops GDI entirely. The cursor is now composited in software using DDA's own PointerPosition + GetFramePointerShape, in three blend modes (color alpha-blend, masked-color XOR, monochrome AND/XOR for the inverting I-beam case). No GDI, no MISC_GDI_COMPATIBLE flag, no GPU sync barrier. cursor=True on mover.py now runs at ~117 fps with tight error bars (115-119 across 5 runs), basically indistinguishable from cursor=False (~125 fps under the same noise).

This was the single biggest correctness/perf issue in the v0.0.5 line. If you've been using cursor=True and seeing dropped frames under windowed apps, v0.0.6 is the upgrade you want.

The cursor=True cell on mover.py was 47 fps in v0.0.5. The GDI cursor compositing path (DrawIconEx on the GDI-compatible BGRA texture) used to sync with the GPU every frame; that sync stalled under DWM contention. v0.0.6 removes that entirely.

Why this is faster

Every existing PyPI screen-cap library does the DDA loop FROM PYTHON. They acquire each frame through comtypes proxies, allocate a numpy array per call, do format conversion through cv2.cvtColor (bettercam pulls OpenCV in just for that), and hold the GIL the whole time. The native rate the OS can give you (one frame per compositor tick) gets eaten by all of that.

rustcam does the entire AcquireNextFrame -> CopyResource -> Map -> RowPitch-aware memcpy in a single Rust function call, releases the GIL around it, and reuses the same BGRA + staging textures across calls. Format conversion (BGR / RGB / RGBA / grayscale) is a tight scalar Rust loop that LLVM auto-vectorizes, no OpenCV dependency. Theres nothing clever, its just doing the same DXGI calls bettercam does without the per-frame Python overhead.

Additions vs bettercam:

• proper cursor compositing via IDXGISurface1::GetDC + DrawIconEx(DI_NORMAL), which handles the inverting I-beam over text correctly (DrawIconEx does mask + XOR blending natively)
• a region argument that crops on the way out of the staging-texture map (no extra alloc)
• five output formats (bgra/bgr/rgba/rgb/gray) with no cv2 dependency
• a paced CFR frames(fps=N) iterator that yields (ndarray, slot_wallclock_ts) for video recording, slot-clock pacing in native Rust, no Python-side timer drift
• a start()/stop()/get_latest_frame() background-thread mode (bettercam-parity API), Rust capture loop, mailbox, blocking pull
• grab_gpu() returns a GpuTexture wrapping a shared NT handle around a BGRA D3D11 texture, zero CPU readback, downstream code (CUDA, Vulkan, custom D3D11) can open the handle on its own device
• a context manager so with rustcam.Capturer(...) as cap: releases COM state on exit
• structured exceptions (AccessLost, DeviceError, DuplicationError, CaptureTimeout, CaptureError) carrying the underlying HRESULT

API

import rustcam

cap = rustcam.Capturer(
    output=0,            # IDXGIOutput index, 0 = primary on single-GPU systems
    cursor=True,         # composite the OS cursor into each captured frame
    region=None,         # persistent (l, t, r, b) crop; None = full output
    device=0,            # IDXGIAdapter index, 0 = first adapter
)

# state
cap.width, cap.height            # output resolution
cap.region                       # current persistent region (full if None)
cap.output_idx, cap.device_idx
cap.cursor, cap.format, cap.rotation
cap.is_capturing                 # True between start() and stop(), or during frames() iteration

# one-shot capture
frame = cap.grab(
    timeout_ms=1000,                  # wait up to this long; 0 = poll
    fmt="bgra",                       # bgra / bgr / rgba / rgb / gray
    region=None,                      # per-call crop, doesn't mutate cap.region
)
# returns numpy ndarray (H, W, C) uint8, or None on DXGI_ERROR_WAIT_TIMEOUT

# background capture (bettercam-parity)
cap.start(target_fps=60, video_mode=True)
frame = cap.get_latest_frame(timeout_ms=500)   # blocks until new frame; raises CaptureTimeout on deadline
cap.stop()

# paced CFR stream — yields (ndarray, slot_wallclock_seconds), exact 1/fps spacing
for frame, ts in cap.frames(fps=60, fmt="bgr"):
    encoder.write(frame, pts=ts)
    if ts > 10.0:
        break

# zero-copy GPU handle (consumer opens with OpenSharedResource1 on its own D3D11 device)
tex = cap.grab_gpu(timeout_ms=200)
if tex is not None:
    print(tex.width, tex.height, hex(tex.shared_handle), tex.luid)
    tex.close()    # release the duplicated NT handle

# context manager
with rustcam.Capturer(output=0) as cap:
    frame = cap.grab()

# module helpers
rustcam.list_outputs()               # list of dicts (one per output across all adapters)
rustcam.device_info()                # bettercam-style multi-line string
rustcam.output_info()                # same

Exceptions (all subclasses of rustcam.CaptureError):

• CaptureError - base; catches every DXGI-origin failure
• DeviceError - device removed / reset
• DuplicationError - DuplicateOutput failed (often: another process already capturing this output)
• AccessLost - exclusive fullscreen took over the display; rustcam retries duplication once internally
• CaptureTimeout - raised by the streaming APIs (get_latest_frame, frames) when their deadline expires; grab() returns None on timeout instead

Each carries a .hresult attribute with the raw HRESULT when relevant.

Compatibility notes

A Capturer is bound to the OS thread that created it. Use one per thread. The Rust extension is #[pyclass(unsendable)], so passing a Capturer between threads raises a RuntimeError.

grab(), start(), frames(), and grab_gpu() are mutually exclusive while a long-running operation is active. While the Capturer is in background mode (start()) or iterating frames, calling grab() or grab_gpu() raises RuntimeError. Call stop() (or close the frames() iterator) first.

The first DDA frame after construction is sometimes black. rustcam discards two warmup frames internally so the first user-visible grab() returns real content.

DDA cant see HDCP-protected content (Netflix, Disney+, etc) - that's the DRM working as designed, and you get a black texture. UWP apps with the protected-content flag set behave the same way. There is no way around this without going through different APIs (WGC + ContentDeliveryManager) which are out of scope here.

grab_gpu() returns a shared NT handle. The consumer opens it on its own D3D11 device via OpenSharedResource1. v0.0.3 ships without a strict keyed-mutex protocol on the shared texture; the consumer should copy the texture into its own resource before issuing GPU work that depends on the content. A future release will add an opt-in strict mode.

Future work

• Optional strict keyed-mutex mode for grab_gpu() so a single-writer single-reader pipeline can rely on producer/consumer ordering.
• WGC (Windows.Graphics.Capture) backend as a fallback for per-window capture and HDR sources where DDA can't help.
• 10-bit / HDR backbuffer support.
• ARM64 Windows wheels.

Shared dda_capture crate

The DDA-specific bits (cursor compositor, region/crop, error type) live in a shared Rust crate at github.com/zen-ham/dda_capture so this package and zentape (a native NV12 video encoder that uses the same DDA capture path) can share one implementation. The cursor=True fix in particular was the kind of subtle bug nobody wants to debug twice — having it in one place means a fix to rustcam ports straight to zentape and vice versa. The crate is a normal cargo git dep, no path tricks needed.

License

MIT. See LICENSE.