Video Depth Anything: Consistent Depth Estimation for Super-Long Videos
Paper • 2501.12375 • Published • 22
Video Depth Anything (Small) — ONNX for WebGPU
A mirror of the browser-ready ONNX export used by Lowkey Studio's depth layers. It runs on onnxruntime-web's WebGPU execution provider.
| File | Input → output |
|---|---|
vda_vits_encoder_518x910_fp16.onnx |
pixels [1, 3, 518, 910] float32 RGB, ImageNet mean/std → feat1..feat4 [1, 2405, 384] |
vda_vits_head_T32_37x65_fp16.onnx |
feat1..feat4 [32, 2405, 384] → depth [32, 1, 518, 910] (relative disparity, near = larger) |
The weights are fp16 and the graph I/O is fp32. Shapes are static (opset 17). The graph is split at the ViT-token boundary, so per-frame encoder features can be reused across overlapping 32-frame windows.
Attribution
- Model and architecture: Video Depth Anything by ByteDance (CVPR 2025), arXiv:2501.12375, https://github.com/DepthAnything/Video-Depth-Anything. Only the Small variant is Apache-2.0; the Base and Large variants are CC-BY-NC and are not included.
- ONNX conversion: TheoreticallyTim/theoreticallypose-vda. Files are mirrored unmodified (sha256 below). No retraining or fine-tuning.
9e05204252f8cdd39e3151541ebded90aa07fe1026cb03d0996748f518e56512 vda_vits_encoder_518x910_fp16.onnx
8037ae4cb8f6b775599fae31badb4abd06ab4e5f9c26b86e70d21c7f1e3a4f7a vda_vits_head_T32_37x65_fp16.onnx
License: Apache-2.0 (see LICENSE).
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support