---
license: apache-2.0
library_name: onnx
pipeline_tag: depth-estimation
tags:
  - depth-anything
  - onnx
  - depth-estimation
  - visionserve
base_model: depth-anything/Depth-Anything-V2-Small-hf
---

# Depth Anything V2 Small — ONNX with dynamic input size

ONNX export of [`depth-anything/Depth-Anything-V2-Small-hf`](https://huggingface.co/depth-anything/Depth-Anything-V2-Small-hf)
(Apache-2.0) for [VisionServe](https://hub.docker.com/r/mtbui2010/visionserve):

```bash
visionserve pull depth-anything-v2
```

Only the **Small** variant is published: Depth Anything V2 Base and Large are CC-BY-NC-4.0.

## Why another export

The input is `[1, 3, H, W]` with **dynamic H and W**, so the image can be fed at the size the
reference `DPTImageProcessor` uses (keep the aspect ratio, scale by whichever of 518/W, 518/H is
closer to 1, each side rounded to a multiple of 14, bicubic). For example, an 848×480 photo is
fed at 910×518 instead of being squashed to 518×518.

A plain dynamic-axes export of this model is wrong. The legacy tracer bakes the 518×518 output
size into the head (`int(patch_h * 14)` in `modeling_depth_anything.py`), so every input returns
a 518×518 map. This export keeps that size traced, and it is checked against PyTorch at three
input shapes (518×518, 910×518 and 518×686).

## I/O contract

| | name | shape | |
|---|---|---|---|
| input | `pixel_values` | `[1, 3, H, W]` f32 | H, W multiples of 14; /255, ImageNet mean/std |
| output | `predicted_depth` | `[1, H, W]` f32 | relative inverse depth (disparity), unnormalised |

## Verification (VisionServe converter, 7 non-square photos)

| check | result |
|---|---|
| ONNX vs PyTorch, 3 input shapes | max \|Δ\|/scale 1.05e-05 |
| served preprocessing vs `DPTImageProcessor` | mean 0.25 grey levels, same shapes on 7/7 images |
| served depth map vs `predicted_depth` | Pearson r min 0.9999, mean 1.000 |

## License

Apache-2.0, inherited from `depth-anything/Depth-Anything-V2-Small-hf`. The weights are not modified.