Start with editing priors
Step1X-Edit supplies an Image-to-Image foundation that preserves scene structure. Depth is regressed in one deterministic step.
STRUCTURE-PRESERVING I2I ACM MULTIMEDIA 2026 · RIO DE JANEIRO
OpenDepth.
Seeing beyond
appearances.
Repurposing Image Editing Priors for Robust Open-World Monocular Depth Estimation
The world changes its appearance.
Its geometry shouldn't have to.
01 / OPEN-WORLD PERCEPTION
GEOMETRY
OBSERVATION+ Same scene.
A deeper understanding.
A deeper understanding.
RGB → RELATIVE DEPTHREAL MODEL OUTPUT
79Ktraining samples
1 stepdeterministic inference
Zero-shotopen-world evaluation
I2I priorsstructure-preserving foundation
01 / THE IDEA
Appearance is variable.
Structure is the constant.
Rain, fog and darkness change how a scene looks, but not how it is built. OpenDepth turns this observation into a principle for robust depth estimation.
We repurpose the structure-preserving priors of an image editing model and introduce intra-domain awareness: mapping different appearances of the same geometry into a unified latent space. Shallow feature alignment and joint RGB reconstruction help separate environmental style from underlying structure.
Abstract
Recently, the integration of generative priors has revolutionized Monocular Depth Estimation (MDE), achieving state-of-the-art performance with remarkably high data efficiency. However, although methods based on Text-to-Image (T2I) models excel in general scenes, their efficacy degrades sharply when confronted with open-world environmental perturbations, such as illumination variations and adverse weather. Since T2I priors are optimized for semantic synthesis rather than preserving pixel-level structure, they struggle to decouple geometry from appearance under severe visual shifts, easily triggering structural hallucinations. Therefore, our core insight is to model open-world depth estimation as consistent geometric inference across multi-style images, fully leveraging the powerful structure-preserving editing priors of Image-to-Image (I2I) models. Based on this, we propose OpenDepth, a robust open-world MDE framework that introduces “Intra-Domain Awareness”—the capability to map diverse stylistic inputs sharing identical geometry into a unified latent space. To achieve this, we first align shallow features within the Diffusion Transformer (DiT) to anchor degraded observations to unified geometric representations. Additionally, we introduce a joint RGB latent reconstruction task, requiring the model to recover clean RGB information from degraded inputs. This multi-stage strategy effectively decouples environmental style interference, compelling the model to focus strictly on underlying geometric recovery. Extensive evaluations show that OpenDepth sets new SOTA records for robust zero-shot depth estimation across standard and extreme weather benchmarks, demonstrating exceptional structural integrity and adaptability.
02 / THE ROBUSTNESS LAB
Change the conditions.
Keep the geometry.
Explore actual OpenDepth predictions across visual shifts.
One scene. Six appearances. A consistent sense of depth.
‹ ›
RGB INPUTOPENDEPTHRELATIVE DEPTH → POINT CLOUDDrag to rotate · illustrative reconstruction
KITTI · synthetic visual perturbations · epoch016 disparity checkpoint · 10-step predictions
NearFar · relative scale
Illustrative 3D lifting using a nominal pinhole camera and per-image relative depth. This view is not a metric reconstruction.
03 / BEYOND THE LAB
The real world
doesn't wait for clear skies.
Nighttime traffic. Wet roads. Low visibility.
Zero-shot predictions in real adverse conditions.
RGB input above · OpenDepth prediction below. Colors visualize relative depth, normalized independently for each image. Visualizations use the epoch016 disparity checkpoint with 10 inference steps.
04 / UNDER THE SURFACE
Different styles.
One geometric language.
An image editing prior provides the foundation.
Two complementary constraints make it robust.
Align shallow features
Clean and degraded observations share geometry. Early DiT feature alignment anchors both to a common representation.
LATENT FEATURE ALIGNMENTReconstruct clean RGB
A joint RGB latent reconstruction task guides perturbed features toward the clean domain while predicting depth.
UNIFIED RGB RECONSTRUCTIONTHE OPENDEPTH FRAMEWORKView full size ↗
A SHARED REPRESENTATION
Less style interference.
More geometric consistency.
Without reconstruction, deep features scatter across environmental styles. With unified RGB reconstruction, perturbation clusters align within a shared domain.
t-SNE analysis · Figure 4 in the paper
05 / THE EVIDENCE
Robustness,
measured.
Zero-shot relative depth evaluation.
Lower absolute relative error is better.
Real-world nighttime and adverse weather benchmarks · Paper Table 2
OpenDepth DepthAnything-AC DepthAnything V2
AbsRel (%) as reported in the paper. Comparisons follow the paper's evaluation protocol; they are not a live leaderboard.
79K training samples.
Image editing priors make data-efficient, robust geometric perception possible.
06 / BUILD ON THIS WORK
A new perspective
on depth.
If OpenDepth supports your research,
please consider citing our paper.
@inproceedings{zhang2026opendepth,
title = {OpenDepth: Repurposing Image Editing Priors
for Robust Open-World Monocular
Depth Estimation},
author = {Zhang, Mengke and Yang, Tianyi and
Li, Xinzhuo and Ming, Anlong},
booktitle = {Proceedings of the 34th ACM International
Conference on Multimedia},
year = {2026},
doi = {10.1145/3767308.3836413}
}