Iris: Bringing Real-World Priors into Diffusion Model for Monocular Depth Estimation

TL;DR

Iris is a novel deterministic framework for monocular depth estimation that effectively integrates real-world priors into diffusion models, enhancing detail preservation and generalization from synthetic to real scenes.

Contribution

The paper introduces a two-stage Priors-to-Geometry schedule with spectral gating techniques, enabling better domain transfer and detail retention in monocular depth estimation.

Findings

Significant improvement in depth estimation accuracy.

Strong generalization to real-world scenes.

Efficient with limited training data.

Abstract

In this paper, we propose \textbf{Iris}, a deterministic framework for Monocular Depth Estimation (MDE) that integrates real-world priors into the diffusion model. Conventional feed-forward methods rely on massive training data, yet still miss details. Previous diffusion-based methods leverage rich generative priors yet struggle with synthetic-to-real domain transfer. Iris, in contrast, preserves fine details, generalizes strongly from synthetic to real scenes, and remains efficient with limited training data. To this end, we introduce a two-stage Priors-to-Geometry Deterministic (PGD) schedule: the prior stage uses Spectral-Gated Distillation (SGD) to transfer low-frequency real priors while leaving high-frequency details unconstrained, and the geometry stage applies Spectral-Gated Consistency (SGC) to enforce high-frequency fidelity while refining with synthetic ground truth. The two stages share weights and are executed with a high-to-low timestep schedule. Extensive experimental results confirm that Iris achieves significant improvements in MDE performance with strong in-the-wild generalization.