ZipDepth — Bringing Lightweight Zero-Shot Monocular Depth Anywhere, on Any Device

Efficiency

Accuracy per joule, measured — not assumed.

Fig. 2

Mean AbsRel vs. energy per frame on NVIDIA Jetson Orin NX (15 W); bubble size encodes FPS. All measurements at 384×384 resolution, vanilla FP32 without optimizations, for a clean, reproducible baseline. For the full latency/FPS comparison across backends and platforms, see Deployment and the supplementary material.

397mJ energy per frame, Jetson Orin NX 15 W · FP32 · 384×384

<2.5J embedded-regime energy budget

9 hardware platforms profiled end-to-end

Method

Two streams, one constraint: every operator must export cleanly.

Given an RGB image, ZipDepth produces a full-resolution affine-invariant inverse-depth map through a four-stage hierarchical encoder built on reparameterizable convolutional blocks — fused into plain convolutions at inference, with zero custom operators — and a streamlined decoder with hardware-adaptive convex upsampling that preserves depth boundaries at full resolution. After Conv–BN fusion, the network totals just 6.1M parameters.

Fig. 3

STAGE Encode

Reparameterizable encoder

Four stages at strides {4, 8, 16, 32}. Dilated depthwise context and strip-pooling attention at Stage 2; SE and global-context attention at Stage 3; SPPF with bidirectional cross-scale refinement.

STAGE Decode

Progressive decoder

Coarse-to-fine fusion via grouped 1×1 convolutions down to stride 4, plus a dedicated half-resolution skip from the split stem that recovers fine spatial detail.

STAGE Upsample

Hardware-adaptive upsampling

Two interchangeable full-resolution paths chosen before training: convex upsampling for GPU/TensorRT, and a learned nearest–bilinear blend for mobile NPU/DSP.

Training follows the scale-and-shift-invariant objective of MiDaS / Depth Anything v2, distilled from Depth Anything v2-Large over 14.1M images spanning 17 heterogeneous domains — trained end-to-end on just two NVIDIA RTX 3090 GPUs over three days.

Results

Zero-shot, across five real-world benchmarks.

We compare against large pretrained foundation models and lightweight architectures retrained on the same multi-domain training set, on NYUv2, KITTI, ETH3D, ScanNet, and DIODE.

Table 2

Among lightweight models, ZipDepth ranks first or second on every benchmark, with the best AbsRel on NYUv2 (8.4) and ScanNet (8.8). At 6.1M parameters and 3.0 GFLOPs, it reaches 34.4 FPS at ~397 mJ/frame on a 15 W Jetson Orin NX.

Qualitative comparison

Drag the handle on any panel (or focus it and use ← →) to reveal ZipDepth's prediction beneath the input. Zero-shot, captured live on iPhone 12.

Citation

@inproceedings{tosi2026zipdepth,
  title     = {{ZipDepth}: Bringing Lightweight Zero-Shot Monocular Depth Anywhere, on Any Device},
  author    = {Tosi, Fabio and Bartolomei, Luca and Poggi, Matteo and Mattoccia, Stefano},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}

Lightweight Zero-Shot Monocular Depth, Anywhere, on Any Device.