Search papers, labs, and topics across Lattice.
The paper introduces Iris, a deterministic diffusion-based framework for monocular depth estimation (MDE) that incorporates real-world priors to improve generalization and detail preservation. Iris employs a two-stage Priors-to-Geometry Deterministic (PGD) schedule, using Spectral-Gated Distillation (SGD) to transfer low-frequency real priors and Spectral-Gated Consistency (SGC) to enforce high-frequency fidelity. Experiments demonstrate that Iris achieves significant MDE performance improvements and strong in-the-wild generalization with limited training data.
By injecting real-world priors into a diffusion model, Iris achieves state-of-the-art monocular depth estimation with significantly improved generalization and detail, even with limited training data.
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.
