ZipDepth ONNX

FP32 ONNX conversion of Fabio Tosi et al.'s ZipDepth, with the original learned attention and convex upsampling preserved. No quantization or retraining is applied.

License and provenance

The upstream repository distributes its code and checkpoints under its MIT LICENSE; the original copyright and permission notice is included in LICENSE. This conversion is not an official upstream release. ZipDepth was distilled from Depth Anything V2 Large; this card does not relicense the teacher or its original datasets, which are not included here.

Input and output

Input image is FP32 RGB divided by 255, NCHW [1,3,768,768]. Preserve the source aspect ratio when resizing and pad the remaining area with black. Ignore padded regions when displaying or exporting scene points. Output inverse_depth is FP32 [1,1,768,768], affine-invariant inverse depth: larger means nearer. It is not a distance in meters. Calibrated distance requires a scale and shift in inverse-depth space. A relative preview may use z = scale / (raw + shift) with positive denominators, while retaining the raw output unchanged.

Validation

The JSON manifest contains checkpoint and ONNX SHA-256 hashes and numerical comparisons with the original PyTorch model on a public image, RGB noise and black input. Browser latency and WebGPU operator support must be measured separately; CPU conversion checks are not browser benchmarks.

Build

Source: https://github.com/DaisukeDaisuke/ZipDepth Every successful build updates this same repository. Use a Hugging Face commit revision to pin an immutable model download. Source commit: 5d7918b55b9204ee24fc76f0c16778d612402d3e ONNX SHA-256: 7926169714cc7e310408f4640f650db3253be1453ad185384bddbea13339795b

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support