ApDepth: Aiming for Precise Monocular Depth Estimation Based on Diffusion Models

Jiawei Wang^a,c,d,† Mingbo Lei^a,c, Liang Yao^a,c Jiayi Li^a,c, Tianyu Lu^a,c, and Yibo Chen^b†,

^aSchool of Computer Science and Engineering, Shenyang Jianzhu University
^bSchool of Computer Science and Technology, Zhejiang Sci-Tech University
^cShenyang Branch of the National Special Computer Engineering Technology Research Center
^dInfiniFlow (Shanghai) Information Technology Co., Ltd.
† These authors contributed equally to this work.

Abstract

This paper presents ApDepth, a novel single-step diffusion framework for monocular depth estimation that enables fast inference while preserving fine-grained edge details. Current diffusion-based models achieve impressive generalization through iterative denoising. However, they face a critical trade-off: multi-step inference incurs prohibitive latency, whereas compressing the process into a single deterministic step can severely degrade local structural boundaries. Our model addresses this challenge by integrating a pre-trained, data-driven MDE model that provides strong semantic priors and maintains robust global structure despite the compressed diffusion process. To mitigate the inherent edge degradation, we introduce a coarse-to-fine two-stage training paradigm with two novel designs: 1) Feature Alignment via Cosine Similarity, which enriches the U-Net's structural understanding by explicitly aligning its intermediate features with an auxiliary encoder; and 2) an Iteration-Based Loss Scheduling Strategy, which initially employs spatial-domain constraints (latent MSE and pixel L₁ losses) to establish an accurate global metric scale and subsequently transitions to a novel Frequency-Domain (FFT) Loss in the final training phase to explicitly recover sharp high-frequency edges. By effectively balancing inference efficiency and detail preservation, ApDepth significantly accelerates inference and achieves performance competitive with or superior to existing methods across multiple benchmarks. Code is available at https://haruko386.github.io/research/.

How it works

Fine-tuning protocol

In the first stage, we learn robust and generalizable representations by aligning U-Net features with high-quality semantic features from DINOv2, emphasizing global context while avoiding overfitting to low-level textures. Building on these semantic priors, the second stage focuses on depth estimation and detail refinement: it first captures global structure and scale through joint optimization, and then enhances fine details and sharp discontinuities via frequency-domain supervision, achieving both structural accuracy and high-fidelity reconstruction.

Single-Step Inference

During inference, the input RGB image is first processed by a pretrained depth estimation model to generate a coarse predicted depth map. Both the RGB image and the predicted depth map are encoded into the latent space using the frozen VAE encoder to obtain the corresponding latent representations. These two latents are concatenated and fed into the Latent Diffusion U-Net, which performs only a single denoising step to generate the predicted latent representation. Finally, the latent decoder reconstructs the refined depth map from the predicted latent, producing the final output depth.

We train our model on a single NVIDIA RTX 6000 Ada Generation GPU , 20000 iters * 32 batchs for stage one and 1000 iters * 32 batchs for stage two. It takes around 6 days to complete the whole training process. We test our model on a single NVIDIA Geforce RTX 5090 GPU , result are available below.

Compared with Other Model-Driven Methods

Zero-shot monocular depth estimation benchmark comparison.
Type	Method	Training Data	NYUv2		KITTI		ETH3D		ScanNet		DIODE
Type	Method	Training Data	AbsRel↓	δ₁↑	AbsRel↓	δ₁↑	AbsRel↓	δ₁↑	AbsRel↓	δ₁↑	AbsRel↓	δ₁↑
Discriminative	DiverseDepth	320K	11.7	87.5	19.0	70.4	22.8	69.4	10.9	88.2	–	–
	MiDaS	2M	11.1	88.5	23.6	63.0	18.4	75.2	12.1	84.6	–	–
	LeReS	354K	9.0	91.6	14.9	78.4	17.1	77.7	9.1	91.7	–	–
	Omnidata	12.2M	7.4	94.5	14.9	83.5	16.6	77.8	7.5	93.6	–	–
	HDN	300K	6.9	94.8	11.5	86.7	12.1	83.3	8.0	93.9	–	–
	DPT	1.4M	9.8	90.3	10.0	90.1	7.8	94.6	8.2	93.4	–	–
	Depth Anything V2	63.5M	4.3	98.0	8.0	94.6	6.2	98.0	4.3	98.1	6.6	95.2
Generative	Marigold	74K	5.5	96.4	9.9	91.6	6.5	96.0	6.4	95.1	10.0	90.7
	GeoWizard	280K	5.2	96.6	9.7	92.1	6.4	96.1	6.1	95.3	12.0	89.8
	DepthFM	74K	6.0	95.5	9.1	90.2	6.5	95.4	6.6	94.9	–	–
	GenPercept	90K	5.6	96.0	9.9	90.4	6.2	95.8	5.6	96.5	–	–
	Lotus	54K	5.3	96.7	8.5	92.8	5.9	97.0	5.9	95.7	9.8	92.4
	DepthMaster	74K	5.0	97.2	8.2	93.7	5.3	97.4	5.5	96.7	–	–
	ApDepth (Ours)	74K	4.5	97.3	8.7	93.7	5.6	96.3	4.5	97.3	8.3	93.2

Citation

Please cite our paper if you find it useful:

@misc{wang2026apdepth, title = {{ApDepth}: Aiming for Precise Monocular Depth Estimation Based on Diffusion Models}, author = {Wang, Jiawei and Lei, Mingbo and Yao, Liang and Li, Jiayi and Lu, Tianyu and Chen, Yibo}, year = {2026}, doi = {10.2139/ssrn.7604741}, url = {https://doi.org/10.2139/ssrn.7604741}, note = {SSRN Preprint} }