VisionDepth3D v4.2.2 - DirectML build General Bug Fixes

VisionDepth3D v4.2.2 - DirectML build General Bug Fixes

This update fixes a major DA-V2 slowdown caused by Original inference resolution being treated as full-resolution model input.

DA-V2 Original mode now keeps the source aspect ratio while using a safer default max side of 518, preventing accidental 1080p/4K model inference and improving practical render speed.

Also included:

• faster DA-V2 adapter path
• less unnecessary resizing
• less duplicate normalization
• reduced GPU-to-CPU transfer size
• clearer model/backend/FP16 diagnostics

Advanced users can still force true original-resolution DA-V2 inference in source code build with:

VD3D_DAV2_DEFAULT_MAX_SIDE=0

Upgrade Instructions

Before installing this update:

1. Back up your weights folder.

2. Back up your presets folder.

3. Run the VisionDepth3D Update Application.

4. Click Uninstall VD3D to remove your current VisionDepth3D installation.

5. In the update app, choose the correct installer build from the dropdown near the top:

   • CUDA for NVIDIA GPU users

   • DirectML for AMD, Intel, or DirectML users

6. Download the selected installer.

7. Run the downloaded Inno Setup installer.

8. Finish setup.

9. Open VisionDepth3D.

10. Load your media and start converting.

Build

This is the DirectML build. If you have CUDA, use the CUDA build.

Final Notes

VisionDepth3D v4.2.2 is was a general bug patch for depth engine and depth blender

Thanks to everyone testing, reporting bugs, and helping push VisionDepth3D forward.

VisionDepth3D v4.2.2 - CUDA Build General Bug Fixes

VisionDepth3D v4.2.2 - CUDA build General Bug Fixes

This update fixes a major DA-V2 slowdown caused by Original inference resolution being treated as full-resolution model input.

DA-V2 Original mode now keeps the source aspect ratio while using a safer default max side of 518, preventing accidental 1080p/4K model inference and improving practical render speed.

Also included:

• faster DA-V2 adapter path
• less unnecessary resizing
• less duplicate normalization
• reduced GPU-to-CPU transfer size
• clearer model/backend/FP16 diagnostics

Advanced users can still force true original-resolution DA-V2 inference in source code build with:

VD3D_DAV2_DEFAULT_MAX_SIDE=0

Upgrade Instructions

Before installing this update:

1. Back up your weights folder.

2. Back up your presets folder.

3. Run the VisionDepth3D Update Application.

4. Click Uninstall VD3D to remove your current VisionDepth3D installation.

5. In the update app, choose the correct installer build from the dropdown near the top:

   • CUDA for NVIDIA GPU users

   • DirectML for AMD, Intel, or DirectML users

6. Download the selected installer.

7. Run the downloaded Inno Setup installer.

8. Finish setup.

9. Open VisionDepth3D.

10. Load your media and start converting.

Build

This is the CUDA build. If you have AMD/Intel GPU then use the DirectML build.

Final Notes

VisionDepth3D v4.2.2 is was a general bug patch for depth engine and depth blender

Thanks to everyone testing, reporting bugs, and helping push VisionDepth3D forward.