A virtual camera films an AI agent's screen while a tracker locks onto every glyph. Two fframes passes, one SkSL shader, all sound synthesized.
Windows setup
The Windows build uses Vulkan in place of Metal. Scenes, camera math, shader and sound synthesis are unchanged.
Requirements
- 64-bit Windows and a Vulkan-capable GPU with current drivers.
- Rust with the MSVC toolchain.
- Visual Studio Build Tools with Desktop development with C++, including a Windows SDK.
- Python 3 with
numpyandscipy. - LLVM/libclang, available from LLVM releases.
- A shared FFmpeg 9.0 development build, with
bin,includeandlibdirectories. For example, useffmpeg-n9.0-latest-win64-gpl-shared-9.0.zipfrom BtbN/FFmpeg-Builds.
A standalone ffmpeg.exe is insufficient: Rust links against FFmpeg libraries,
and build-time media processing loads the DLLs. See also
fframes' Windows requirements.
Render
In PowerShell, adjust the installation paths to match your machine:
$env:FFMPEG_DIR = "C:\tools\ffmpeg-n9.0-latest-win64-gpl-shared-9.0"
$env:LIBCLANG_PATH = "C:\Program Files\LLVM\bin"
$env:PATH = "$env:FFMPEG_DIR\bin;$env:PATH"
git clone https://github.com/mrsarac/ff-tracking
cd ff-tracking
python -m venv .venv
.\.venv\Scripts\python.exe -m pip install numpy scipy
powershell -NoProfile -ExecutionPolicy Bypass -File tools/render.ps1 -Python .\.venv\Scripts\python.exe
# -> out/tracking.mp4
The execution-policy option applies only to this PowerShell process. Keep FFmpeg's
bin directory on PATH when running the built programs. The script accepts the
Python executable via -Python, stops on failed steps, restores the caller's
directory, and honors Cargo's configured target directory.
After rendering, the usual per-pass tools work:
.\target\release\lens.exe preview
.\target\release\lens.exe strip all -n 12
.\target\release\lens.exe audio analyze
If you configured a custom Cargo target directory, use its release folder instead.
Windows uses the codecs from the shared FFmpeg installation, so the Cargo features that request building codecs from source are enabled only for the original non-Windows configuration. Windows limits the rendering pipeline to two GPU contexts, two encoder workers and a small queue: each worker decodes the 4K first pass, so unrestricted concurrency can exhaust memory, especially on longer edits. The original macOS backend and concurrency setting remain unchanged.