Zero-shot monocular depth in 6.1M parameters — "ZipDepth," lightweight depth for any device (ECCV 2026)

This work (ZipDepth) advances monocular depth estimation — inferring distance from a single image — on two fronts at once: small enough for constrained devices, and robust enough not to break in unseen domains.

1The gap between foundation models and devices

Monocular depth estimation has progressed rapidly through foundation models. Trained at scale, these models generalize robustly zero-shot to environments they have never seen. But their computational demands put them far beyond what embedded systems and mobile devices can run.

Lightweight alternatives exist, yet they have been developed almost exclusively within single-domain, self-supervised paradigms — so under domain shift they fail silently, returning wrong depth without any error.

2An efficient encoder-decoder design

ZipDepth closes this gap with two combined elements. First, an efficient, reparameterizable encoder-decoder: the network trains in an expressive configuration and folds, via equivalent transformation, into a faster form at inference.

Second, large-scale knowledge distillation from a foundation model over a large multi-domain training set — transferring the generalization ability of the foundation model itself into the compact network, rather than relying on single-domain self-supervision.

3Real-time at 6.1M parameters

The result: at just 6.1M parameters, ZipDepth runs at real-time rates from server GPUs down to power-constrained devices. Across five benchmarks it achieves the best trade-off between zero-shot accuracy and deployment efficiency among lightweight models, taking a significant step toward the accuracy of foundation models with 50x more parameters.