FuryTMP/Video-Depth-Anything-L-ONNX-512x288

Video Depth Anything (ONNX) – VisionDepth3D Edition

This is a converted ONNX version of Video Depth Anything (VDA), prepared for use with VisionDepth3D and other ONNX-compatible pipelines.

About

• Original weights: VDA (PyTorch .pth)
• Converted to ONNX format for faster inference and integration
• Tested and optimized for VisionDepth3D v3.5+

Input / Output

• Input resolution is fixed at 512×288
• The ONNX export does not currently support dynamic shapes
• Outputs a single-channel depth map normalized to [0..1]

Use Cases

• Depth map generation for video frames
• 2D-to-3D conversion inside VisionDepth3D
• Preprocessing for VR, AR, and cinematic pipelines

Notes

• This conversion is for compatibility; accuracy is consistent with the original model
• For real-time use, GPU acceleration via CUDA is strongly recommended
• Works directly in VisionDepth3D’s ONNX inference pipeline

How To Use

• Download the VideoDepthAnything folder in files tab
• Put the folder into the weights directory of VisionDepth3D
• In VD3D Depth Estimation Tab in model dropdown scroll to bottom of the list and choose the (ONNX) model not the (Local) model
• Set input resolution to 512x288 (wouldnt convert for dynamic input resolution)

Citation

Please cite the original Video Depth Anything project if you use this model. This ONNX export is provided only for compatibility and workflow integration.