---
license: mit
pipeline_tag: depth-estimation
tags:
  - monocular-depth
  - relative-depth
  - zero-shot
  - lightweight
---

# ZipDepth (base) — pretrained checkpoints

Mirror of the official ZipDepth release checkpoints for use with
`hf_hub_download` (e.g. the `monopriors` package in rerun-io/examples-monorepo).

**ZipDepth: Bringing Lightweight Zero-Shot Monocular Depth Anywhere, on Any Device**
Tosi, Bartolomei, Poggi, Mattoccia — ECCV 2026.
Paper: https://arxiv.org/abs/2607.08771 · Code: https://github.com/fabiotosi92/ZipDepth
(pixi fork: https://github.com/pablovela5620/ZipDepth)

| File | Upsampling | Params (fused) | Use |
|------|-----------|----------------|-----|
| `zipdepth_base.pth` | Convex (`torch.nn.Unfold`) | ~6.1 M | GPU / server (`upsample_unfold=True`) |
| `zipdepth_base_npu.pth` | Convex, unfold-free | ~6.1 M | NPU / mobile / ONNX export (`upsample_unfold=False`) |

Both share identical encoder/decoder weights; only the upsampling head differs.
Output is relative inverse depth (disparity up to scale). Trained at 384×384.

```python
from huggingface_hub import hf_hub_download
ckpt = hf_hub_download("pablovela5620/zipdepth", "zipdepth_base.pth")
```

License: MIT (see `LICENSE`, © the ZipDepth authors).