[UE 5.8.2] MetaHuman Markerless Motion Capture Stage 4 causes GPU reset and black screens on RTX 5070

Hello, I am experiencing a repeatable GPU/display crash when using the experimental MetaHuman Animator Markerless Motion Capture plugin in Unreal Engine 5.8.2.

System specifications:

  • Windows 11
  • Unreal Engine 5.8.2, build 56702186
  • MetaHuman Animator Markerless Motion Capture plugin 1.0.0 Beta
  • NVIDIA RTX 5070, 12 GB VRAM
  • AMD Ryzen 9 9950X3D
  • 32 GB system RAM
  • NVIDIA Studio Driver 616.56
  • DirectX 12

Source video:

  • H.264 MP4
  • 1080 × 1920
  • 30 FPS
  • Approximately 7 seconds

Problem:

Processing completes Stages 1–3 normally. As soon as Stage 4 begins, all connected monitors lose signal and become completely black. The computer appears to remain powered on and working, but the display never recovers. I eventually have to force the computer to shut down.

This has happened repeatedly, including with a very short video.

The Unreal log reports ONNX Runtime/DirectML errors during Stage 4, followed by a GPU device reset/removal:

DXGI_ERROR_DEVICE_RESET
DXGI_ERROR_DEVICE_REMOVED
HRESULT: 887A0005
onnxruntime.dll
DmlExecutionProvider
ExecutionProvider.cpp(952)

This does not appear to be a normal VRAM out-of-memory error. At the time of the failure, approximately 6.9 GB of VRAM was being used, with an approximately 11.2 GB budget. The Unreal crash information also reported bIsOOM=0.

Troubleshooting already performed:

  • Clean installation of NVIDIA Studio Driver 616.56
  • The problem also occurred with NVIDIA driver 595.76
  • Restarted Windows after driver installation
  • Removed custom TdrDelay and TdrDdiDelay registry values
  • Disabled ray tracing
  • Turned off the 4K monitor and tested using only a Full HD monitor
  • Tested body tracking with face tracking disabled
  • Tested only 30 frames
  • Disabled the preview pass
  • Used a fresh temporary performance asset
  • Confirmed DirectX 12 is enabled
  • Confirmed the problem still happens with the short 7-second clip

The 30-frame body-only test produced the same DirectML error at Stage 4 and reset the display driver.

Comparison:

The markerless body-capture workflow completed successfully on another computer using:

  • Unreal Engine 5.8.0
  • RTX 5090
  • Ryzen processor
  • 128 GB RAM
    (later i tried to install 5.8.0 to main PC rtx 5070 but UE doesnt let to install it as i had latest 5.8.2)

Therefore, I suspect a compatibility issue involving one of the following:

  • MetaHuman Animator Markerless Motion Capture plugin and UE 5.8.2
  • ONNX Runtime/DirectML and the RTX 5070
  • A regression between UE 5.8.0 and UE 5.8.2
  • Different memory requirements or plugin builds

Has anyone reproduced this on an RTX 5070 with UE 5.8.2?

Is there a known issue or workaround for the Stage 4 DirectML error 887A0005? Is it possible to test the UE 5.8.0 version of the plugin without compiling the complete engine from source?

I can provide the complete Unreal log if required.