reliquary-depth-anything-v2-small

Depth Anything V2 Small (ViT-S, ~25M params) ONNX export. Powers Reliquary's Heightmap Generator with relative depth estimation.

Source

ONNX re-export of the upstream model — original credits to the authors:

• Repo: https://github.com/DepthAnything/Depth-Anything-V2
• Paper: "Depth Anything V2" — Yang, Kang, Huang, Xu, Feng, Zhao, NeurIPS 2024. Project page: https://depth-anything-v2.github.io/
• Pretrained weights origin: HuggingFace depth-anything/Depth-Anything-V2-Small

License

Inherited from upstream — the Small variant is explicitly Apache-2.0 in the Depth Anything V2 repo. Free for commercial and non-commercial use with attribution.

Usage in Reliquary

This model is consumed by Reliquary, a browser-based PBR texture editor. The ONNX file is loaded via ONNX Runtime Web in the browser — no Python or server-side inference.

• Runtime integration point: src/lib/ai/depthAnything.ts (variant picker)
• Input shape / format: [N, 3, H, W] — RGB normalised. H, W must be multiples of 14 (ViT patch size). Reliquary pads input to the next multiple of 14 before feeding the session.
• Output shape / format: [N, 1, H, W] — relative (disparity-like) depth, larger = closer. Normalise per-image to [0, 1] for heightmap use.

Files

• depth_anything_v2_small.onnx — ONNX graph (1.0 MB)
• depth_anything_v2_small.onnx.data — external weights (94.4 MB, present because the model exceeds the ONNX protobuf size limit)
• README.md — this file

Disclaimer

This is a re-export from the original PyTorch checkpoint to a browser-deployable ONNX format with dynamic input axes. The model weights, architecture, and training methodology are entirely the work of the original authors cited above. This repo is a hosting convenience for Reliquary's browser-side ML — it does NOT modify the model.

For the full conversion pipeline (including version-compat shims for PyTorch 2.6+, opset selection, dynamic-axes details), see Reliquary's PYTHON_CONVERSIONS_GUIDE.md.