Deep Live Cam Studio 2.2.0
Deep Live Cam Studio 2.2.0 Windows Release
This release adds a verified DirectML build for AMD and Intel GPUs and fixes a
file Preview/Render deadlock that could also affect CUDA systems.
Downloads
- NVIDIA/CUDA installer:
DeepLiveCamStudio-2.2.0-x64-setup.exe - Installer SHA-256:
55B83C48BDCDF818E3F051AD2896A1495AED1DEA2E676BD3F8E071A7DC45916E - AMD/Intel DirectML portable ZIP:
DeepLiveCamStudio-2.2.0-DirectML-x64-portable.zip - DirectML portable SHA-256:
177AD450FB2BBBE34B5B9F0BACFB77D7541BFD16201524B6A4324A02E57A32FB - Corresponding source archive:
DeepLiveCamStudio-2.2.0-source-0e2617e0c944.zip - Source archive SHA-256:
A4BB8F774FEA453AE9077C8F7E3D004B5297DEBF9CE9A9FF0C0CAF56417696D5 - Source ref:
0e2617e0c944f911aa26c144878292e0272270c9
Use the CUDA installer for NVIDIA GPUs. Use the DirectML ZIP for AMD or Intel
DirectX 12 GPUs: extract it into a new folder and run
DeepLiveCamStudio.exe. Do not copy one runtime over another.
What Changed
Added
- DirectML GPU acceleration for AMD and Intel GPUs, including strict provider
verification and multi-GPU adapter selection. - A versioned DirectML portable ZIP with a SHA-256 sidecar and release-time
archive validation.
Fixed
- Preview and Start Render can no longer re-enter while the first file
operation is loading models, preventing the observed Not Responding
deadlock. - DirectML face analysis uses the CPU compatibility path while swap and
enhancement inference remain on the GPU, avoiding Radeon multi-session
hangs. - GitHub packaging now preserves dot-prefixed runtime directories, including
the scikit-learn.libsDLL required by the application. - Clean NVIDIA installer builds now include the pinned CUDA 12/cuDNN 9 runtime
DLL set required for an actual ONNX Runtime CUDA session. - Provider diagnostics fail clearly when requested GPU acceleration is not
active instead of silently treating CPU fallback as success.
The final DirectML artifact was validated by the issue reporter on a Radeon
6900 XT for file rendering and Live Output. CUDA rendering and Live Output were
also verified locally on an NVIDIA RTX 4070.
Source and license
Deep-Live-Cam is licensed under AGPL-3.0. Complete corresponding source for the
exact binary release is attached and identified by the source ref above.
The source archive is produced with:
powershell -ExecutionPolicy Bypass -File build\windows\package_source.ps1 -AppVersion 2.2.0 -GitRef 0e2617e0c944f911aa26c144878292e0272270c9This release preserves attribution to the original project:
- Original project: https://github.com/hacksider/Deep-Live-Cam
- License: AGPL-3.0
Models are not bundled
The installer intentionally does not include model/checkpoint files. The
DirectML portable ZIP also excludes them because model licenses and
redistribution rights are separate from the application license.
After launching, select Set Up Models, or run:
DeepLiveCamStudioCLI.exe --download-modelsDownloaded models are stored under:
%LOCALAPPDATA%\DeepLiveCamStudio\models
Requirements and known limitations
ffmpegandffprobemust be installed separately for video workflows.- OBS Studio must be installed separately for OBS Virtual Camera workflows.
- The public files are unsigned and may trigger Microsoft Defender
SmartScreen. - Face-swap and enhancer model families retain their documented
non-commercial or research-use restrictions; no model weights are
redistributed here.
Release evidence
This release candidate is not publish-approved by release notes alone;
publish approval is recorded in RELEASE_VERIFICATION.md,
MANUAL_RELEASE_GATES.md, and the signed gate documents included with the
upload.
Completed local evidence is included in the uploaded release documents:
- Full automated test suite.
- Packaged CUDA and DirectML runtime/provider checks.
- Installer smoke install/uninstall checks.
- Model download/checksum and forbidden-model scans.
- CUDA and DirectML file-render checks.
- Radeon 6900 XT and NVIDIA RTX 4070 Live Output checks.
Remaining publish blockers:
- None when
RELEASE_VERIFICATION.mdsaysReady to publish without remaining manual gates: YES.
Completed manual signoffs include:
- Clean Windows install without admin rights.
- OBS Virtual Camera and DirectML Live Output workflows.
- Authorized legal review for dependency, model-license, and redistribution obligations.