TheoreticallyTim/theoreticallypose-vda

Video Depth Anything (Small) — ONNX for TheoreticallyPose

Browser-ready ONNX export of Video-Depth-Anything-Small (Apache-2.0), used by the "Consistent (VDA)" depth engine in TheoreticallyPose — a single-file browser app that turns a source clip into pose / depth / silhouette control videos for video-to-video AI generation. The app downloads these two files automatically the first time you use the Consistent engine (~59 MB, cached by your browser).

Files

file	what it is	input → output
vda_vits_encoder_518x910_fp16.onnx (45 MB)	per-frame DINOv2-S encoder	[1,3,518,910] normalized RGB → 4 × [1,2405,384] ViT features
vda_vits_head_T32_37x65_fp16.onnx (14 MB)	temporal DPT head, 32-frame window	4 × [32,2405,384] stacked features → [32,1,518,910] depth

Conversion notes: exported with the PyTorch legacy tracer at opset 17, static shapes (518×910, T=32), split at the ViT-token boundary so per-frame encoder features can be cached across overlapping windows; vanilla attention throughout (no xformers); weights converted to fp16 with fp32 graph I/O. Runs on onnxruntime-web's WebGPU execution provider.

Attribution & license

Original model and architecture: Video Depth Anything by ByteDance (repo, paper, CVPR 2025 Highlight). Only the Small variant is redistributed here, under its original Apache-2.0 license (see LICENSE). The Base/Large variants are CC-BY-NC and are not included. This repo contains a format conversion (ONNX/fp16) of the original checkpoint; no weights were retrained or fine-tuned.

@inproceedings{video_depth_anything,
  title={Video Depth Anything: Consistent Depth Estimation for Super-Long Videos},
  author={Chen, Sili and Guo, Hengkai and Zhu, Shengnan and Zhang, Feihu and Huang, Zilong and Feng, Jiashi and Kang, Bingyi},
  booktitle={CVPR},
  year={2025}
}