Recovering detailed geometry from high-resolution images is critical for precise perception of the surroundings and objects. However, existing methods which use latent-space modeling and VAE reconstruction can compromise geometric details. Furthermore, decoding from latent codes introduces substantial inference overhead. To address those issues, we present EagleDepth, an efficient framework for high-resolution monocular depth estimation that combines the geometric priors of latent diffusion with fine-grained pixel-space generation. Our key idea is to retain depth-aware latent representations as guidance while generating the final depth map directly in pixel space. We train the latent and pixel components sequentially: first, we fine-tune a pretrained latent diffusion model using paired RGB--depth supervision; then, we adapt a pretrained pixel diffusion decoder, PiD, to predict depth conditioned on the learned features. Training of the pixel component starts at 1024 resolution and continues across multiple resolutions up to 4K. The latent branch processes resized, lower-resolution RGB images, while the pixel branch generates depth at the target resolution, bypassing the original VAE decoder. This design preserves learned geometric knowledge without requiring the latent backbone to operate at the output resolution. On five commonly used depth estimation datasets and the high-resolution Synth4K dataset, our framework achieves state-of-the-art depth estimation performance, with faster inference and better preservation of fine structures and object boundaries.
Figures & tables
Figure 1: EagleDepth excels at predicting fine-grained depth maps from high-resolution images. Compared with state-of-the-art methods, our model achieves better depth estimation results with faster inference, as shown in the radar chart (quality metrics averaged over the five Synth4K ( Yu et al., 2026a ) subsets; larger radii indicate better performance)
Figure 2: Method overview. (a) An adapted latent predictor, trained with a pixel-space loss, estimates a coarse depth latent from a downsampled RGB image. (b) Conditioned on this latent, the pixel diffusion decoder generates a high-resolution depth map with fine-grained details. (c) A lightweight continuity-promoting module mitigates discontinuity artifacts at patch boundaries.
Figure 3: Qualitative comparison with leading monocular depth estimation methods. Our method preserves fine geometric details and thin structures while maintaining sharp depth boundaries and smooth surfaces, as highlighted by the arrows and zoomed-in regions.
Method
NYUv2
KITTI
ETH3D
ScanNet
DIODE
AbsRel ↓
δ1↑
AbsRel ↓
δ1↑
AbsRel ↓
δ1↑
AbsRel ↓
δ1↑
AbsRel ↓
δ1↑
MiDaS
11.1
88.5
23.6
63.0
18.4
75.2
12.1
84.6
—
—
DA2
4.5
97.9
7.4
94.6
13.1
86.5
6.5
97.2
6.6
95.2
Marigold
5.5
96.4
9.9
91.6
6.5
96.0
6.4
95.1
10.0
90.7
Lotus-2
5.0
97.6
6.6
95.7
5.1
97.5
5.2
97.3
6.1
95.2
DA3
4.1
98.0
5.9
97.2
4.4
97.8
3.9
97.9
5.9
95.1
Table 1: Quantitative results on standard depth benchmarks. Bold and underlined values denote the best and second-best results, respectively. Our method achieves the best results on four of the five benchmarks and remains competitive on KITTI.
Method
Synth4K-1
Synth4K-2
Synth4K-3
Synth4K-4
Synth4K-5
AbsRel ↓
δ1↑
AbsRel ↓
δ1↑
AbsRel ↓
δ1↑
AbsRel ↓
δ1↑
AbsRel ↓
δ1↑
Overall
DA3
38.3
45.7
29.7
51.2
32.2
53.2
23.7
63.9
23.8
58.9
MoGe-2
22.1
67.1
16.6
76.6
18.0
79.3
14.1
82.9
11.0
85.9
PPD
15.9
79.4
11.7
87.3
12.7
87.7
11.9
88.2
9.1
93.1
PointDiT
44.2
42.6
40.4
40.2
47.7
36.6
29.6
55.6
39.1
38.4
URGT
15.0
81.1
10.6
87.5
15.3
84.9
7.4
94.0
7.6
92.6
Table 2: Depth accuracy on the high-resolution Synth4K benchmark, evaluated over the full valid region (Overall) and high-frequency regions. Results for InfiniDepth † are reported with a 504×896 input and a 4K output, following its original paper. All other results are reported with 4K inputs and outputs. Our method demonstrates leading performance at high resolution.
Method
Synth4K-1
Synth4K-2
Synth4K-3
Synth4K-4
Synth4K-5
DA3
8.7
10.8
9.2
11.1
14.1
MoGe-2
15.7
13.1
13.2
14.9
18.2
PPD
15.0
18.4
15.8
18.7
16.1
PointDiT
20.7
21.8
17.4
22.1
20.1
URGT
20.9
16.0
15.0
18.2
21.4
InfiniDepth
15.7
20.0
13.1
23.6
21.7
Table 3: Boundary F1 ↑ (radius = 1) results on Synth4K. Our multi-resolution model achieves the highest scores across all five subsets, demonstrating excellent depth boundary quality.
Figure 4: Ablation study of the proposed continuity-promoting module. The module effectively mitigates discontinuity artifacts between patches.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Dataset
Depth resolution
Selection
Samples
Hypersim
768×1024
AnyDepth filtering
26,695
VKITTI2
375×1242
AnyDepth filtering
12,643
TartanAir
480×640
AnyDepth filtering
186,693
IRS
540×960
AnyDepth filtering
68,211
EDEN
480×640
Official subset
12,520
RobbyReal
480×640 , 720×1280
Every tenth frame
139,575
Appendix
Table 7: Training datasets and sample selection strategies. We report depth resolutions ( H×W ), selection criteria, and the number of selected RGB–depth pairs for each dataset.
Figure 5: Point cloud results of our method.
Component
512×512
1024×1024
2048×2048
2160×3840
Time
Memory
Time
Memory
Time
Memory
Time
Memory
VAE Encoder
16
11.9
83
13.2
424
18.0
1058
24.4
DiT
102
11.8
349
12.4
1733
14.7
4423
17.6
VAE Decoder
34
12.1
159
13.8
732
20.7
1668
29.8
Pixel Diffusion
37
11.8
78
12.4
487
14.8
1298
18.0
CPM
2
11.7
7
11.9
39
13.1
77
14.7
Appendix
Table 9: Component-wise runtime (ms ↓ ) and peak GPU memory (GB ↓ ) analysis. All runtime results are evaluated on an NVIDIA RTX 4090 GPU. CPM denotes the continuity-promoting module.
Figure 6: More visual results (1). Zoom in for better visibility.
Figure 7: More visual results (2). Zoom in for better visibility.
Figure 8: More visual results (3). Zoom in for better visibility.
This work presents Lapis, a linear-attention-based pixel-space generative framework that achieves efficient and high-fidelity depth estimation with one-step diffusion. While generative frameworks have significantly advanced monocular depth estimation with superior detail fidelity, the O(N2) complexity of standard attention and the multi-step denoising process introduce prohibitive computational costs when scaling them to high-resolution image applications. Although linear attention and one-step prediction are intuitively viable, directly applying them leads to poor structural consistency, detail loss, and noise. Lapis rectifies these limitations through a coarse-to-fine hierarchy. Specifically, a Patch-level Consistency Module restores structural coherence by integrating semantic and spatial priors. Subsequently, a Pixel-level Refinement Module recovers sharp geometric boundaries via skip-connection-based pixel correspondence. Furthermore, to mitigate sampling noise inherent in one-step diffusion, we leverage the manifold assumption and adopt a direct x-prediction strategy to target the clean data manifold. Extensive evaluations on multiple benchmarks demonstrate that Lapis consistently achieves state-of-the-art (SOTA) accuracy and boundary sharpness across various resolutions, reducing inference latency by up to 7.6× at 1080P and 10.9× at 1440P resolution compared to previous SOTA generative models.
Bingde Liu, Wu Ran, Jinglei Zhang +2
MoE Key Lab of Artificial Intelligence, Institute of AI, Shanghai Jiao Tong University, Shanghai, China · Zhiyuan College, Shanghai Jiao Tong University, Shanghai, China
Recent monocular depth estimators achieve strong zero-shot generalization, yet often struggle to preserve fine-grained structures and object boundaries. We attribute this limitation to the prevalent combination of large-patch ViT encoders and convolutional decoders, as coarse tokenization can weaken pixel-level cues that upsampling cannot fully recover. To address this issue, we propose PXDepth, a discriminative monocular depth model that separates global context modeling from pixel-level depth prediction. Specifically, a large-patch ViT captures global scene context, while a pixel-space predictor composed of Context-Modulated Pixel Transformer blocks maintains high-resolution spatial representations throughout depth estimation. This design preserves fine structures and sharp boundaries without sacrificing global depth consistency. Across diverse zero-shot benchmarks, PXDepth combines faithful local geometry with competitive global depth accuracy while remaining efficient at inference. Our code and model are available at https://yuanzhy29.github.io/PXDepth-Page/.
Zhiyuan Yuan, Guanying Chen, Lingteng Qiu +3
1Sun Yat-sen University · 3CUHKSZ · 2Shenzhen-FNii
Existing video depth estimation faces a fundamental trade-off: generative models suffer from stochastic geometric hallucinations and scale drift, while discriminative models demand massive labeled datasets to resolve semantic ambiguities. To break this impasse, we present DVD, the first framework to deterministically adapt pre-trained video diffusion models into single-pass depth regressors. Specifically, DVD features three core designs: (i) repurposing the diffusion timestep as a structural anchor to balance global stability with high-frequency details; (ii) latent manifold rectification (LMR) to mitigate regression-induced over-smoothing, enforcing differential constraints to restore sharp boundaries and coherent motion; and (iii) global affine coherence, an inherent property bounding inter-window divergence, which enables seamless long-video inference without requiring complex temporal alignment. Extensive experiments demonstrate that DVD achieves state-of-the-art zero-shot performance across benchmarks. Furthermore, DVD successfully unlocks the profound geometric priors implicit in video foundation models using 163x less task-specific data than leading baselines. Notably, we fully release our pipeline, providing the whole training suite for SOTA video depth estimation to benefit the open-source community.
Hongfei Zhang, Harold Haodong Chen, Chenfei Liao +12