Documentation Index
Fetch the complete documentation index at: https://mintlify.com/hacksider/Deep-Live-Cam/llms.txt
Use this file to discover all available pages before exploring further.
By default, Deep-Live-Cam runs inference on the CPU, which is reliable but slow for real-time use. Enabling a GPU execution provider offloads the ONNX model inference to your hardware accelerator, significantly improving frame rates. The provider you use depends on your GPU: NVIDIA cards use CUDA, Apple Silicon uses CoreML, and AMD/Intel GPUs on Windows can use DirectML or OpenVINO.
Deep-Live-Cam auto-detects the best available provider at startup in this order: cuda → rocm → coreml → dml → cpu. You can override this with the --execution-provider flag.
Choosing a provider
CUDA is the recommended provider for NVIDIA GPUs (Turing architecture or newer for best performance). It requires installing the CUDA Toolkit and cuDNN before updating the Python packages.1. Install system dependencies2. Install Python packagesWith your virtual environment active, install PyTorch with CUDA 12.8 support and the GPU-enabled ONNX Runtime:pip install -U torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128
pip uninstall onnxruntime onnxruntime-gpu
pip install onnxruntime-gpu==1.21.0
python run.py --execution-provider cuda
When CUDA is active, Deep-Live-Cam automatically selects the FP16 inswapper model (inswapper_128_fp16.onnx) when it is present, falling back to the FP32 variant (inswapper_128.onnx) on older GPUs. It also promotes libx264 to h264_nvenc for hardware-accelerated video encoding.
CoreML gives the best performance on Apple Silicon (M1/M2/M3/M4) by routing inference through the Neural Engine and GPU shader cores.Prerequisite: Complete the macOS-specific setup first — install Python 3.11 via Homebrew and create your virtual environment with python3.11 -m venv venv.1. Install the Silicon-specific ONNX Runtimepip uninstall onnxruntime onnxruntime-silicon
pip install onnxruntime-silicon==1.13.1
python3.11 run.py --execution-provider coreml
You must use python3.11 to run the app on macOS, not python or python3, if multiple Python versions are installed. Using the wrong interpreter can cause import errors or load the wrong onnxruntime package.
For older Intel-based Macs or when using the legacy CoreML provider, use the standard onnxruntime-coreml package instead of the Silicon-specific build.1. Install dependenciespip uninstall onnxruntime onnxruntime-coreml
pip install onnxruntime-coreml==1.21.0
python run.py --execution-provider coreml
DirectML works on any DirectX 12-capable GPU on Windows, including AMD and Intel integrated and discrete GPUs. It does not require vendor-specific drivers beyond what Windows already provides.1. Install dependenciespip uninstall onnxruntime onnxruntime-directml
pip install onnxruntime-directml==1.21.0
python run.py --execution-provider directml
DirectML sets the execution thread count to 1 automatically to avoid serialization issues with the DirectML API.
OpenVINO is Intel’s inference optimization toolkit. It works on Intel CPUs, integrated graphics, and discrete Arc GPUs.1. Install dependenciespip uninstall onnxruntime onnxruntime-openvino
pip install onnxruntime-openvino==1.21.0
python run.py --execution-provider openvino
If you do not have a supported GPU, Deep-Live-Cam will run on the CPU using the default onnxruntime package installed with requirements.txt. No additional installation is needed.CPU mode is significantly slower than GPU-accelerated providers and is not suitable for real-time webcam use on most hardware.
| Provider | Hardware | Real-time capable |
|---|
| CUDA | NVIDIA GTX 10xx+ | Yes |
| CoreML (Silicon) | Apple M1–M4 | Yes |
| CoreML (Legacy) | Intel Mac | Limited |
| DirectML | AMD/Intel (Windows) | Varies |
| OpenVINO | Intel | Varies |
| CPU | Any | No |
onnxruntime-gpu==1.21.0. When CUDA is active, Deep-Live-Cam automatically selects the FP16 inswapper model when available (reducing memory bandwidth) and promotes video encoding to h264_nvenc or hevc_nvenc, keeping encoding off the CPU.