Jens-Duttke/ZipDepth-ONNX

ZipDepth - ONNX (FP16)

FP16 ONNX export of ZipDepth (Tosi, Bartolomei, Poggi, Mattoccia, ECCV 2026), a 6.1M-parameter convolutional monocular depth network, built for Oku3D's depth registry and usable by any ONNX Runtime consumer.

file	base checkpoint	size
zipdepth_fp16_opset17.onnx	checkpoints/zipdepth_base.pth at commit 91f3fd2	12 MB

Interface

input   pixel_values     float32  [batch_size, 3, height, width]
output  predicted_depth  float32  [batch_size, 1, height, width]

The input is RGB in [0, 1], not standardized. The network applies the ImageNet mean and standard deviation itself; standardizing beforehand applies them twice.

height and width are free and must be multiples of 32 - the network halves the frame five times. The model was trained at 384; Oku3D runs it at 384 and 512. A consumer can pin the axes by name (batch_size, height, width), which on DirectML runs exactly as fast as a graph exported at that one size.

The output is relative inverse depth, near = high, which needs no inversion for stereo warping. Weights are FP16 while the boundary stays FP32.

Changes against the authors' export

Two operations are replaced by equivalents the exporter can keep free-sized. Both were checked against the original network on random inputs at 384 and 512, with a maximum difference of 0.0:

• F.unfold(depth, 3) in the convex upsampler becomes a 3x3 convolution with nine one-hot kernels in the same order.
• F.adaptive_avg_pool2d in the cross-scale block, whose target is always half the size of its input for a multiple-of-32 frame, becomes F.avg_pool2d(x, 2).

Accuracy and speed

Deviation of this FP16 graph from the FP32 export, as the largest per-pixel difference relative to the depth range, on ONNX Runtime's CPU provider: 0.41 % at 384, 0.51 % at 512.

Inference time with ONNX Runtime 1.24.4 and DirectML, median of 30 runs:

GPU	384	512
NVIDIA GeForce RTX 5070 Ti	0.72 ms	1.18 ms
AMD Radeon integrated graphics (Ryzen 9 9950X)	21.6 ms	36.5 ms

License

MIT, as published by the authors; see LICENSE. According to the paper, ZipDepth was trained on depth maps produced by Depth Anything V2 Large, whose weights are licensed CC BY-NC 4.0.

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}
}