Depth Anything 3: Recovering the Visual Space from Any Views¶

🎯 Pitch¶

Depth Anything 3 (DA3) is a minimal, unified any‑view geometry model that uses a single plain transformer backbone to predict per‑view depth maps and per‑pixel ray maps—enabling direct, spatially consistent 3D reconstruction from any number of images with or without known poses. By reducing targets to a depth–ray representation and leveraging teacher–student training on large synthetic and real data, DA3 achieves state‑of‑the‑art pose and reconstruction accuracy, simplifying deployment across monocular, multi‑view, and rendering tasks and accelerating downstream 3D vision applications (robotics, AR/VR, and novel‑view synthesis).

1. Executive Summary (2-3 sentences)¶

Depth Anything 3 (DA3) is a unified “any-view” geometry model that takes an arbitrary number of images (optionally with known camera poses) and predicts spatially consistent per-view depth plus a per-pixel ray map, enabling direct reconstruction of a shared 3D scene representation (Figure 1; Section 3.1). The core significance is that DA3 aims to replace specialized monocular depth / SfM / MVS-style networks with a minimal transformer-based predictor—using a plain pretrained ViT backbone and a single depth–ray target—while achieving strong pose and reconstruction accuracy on a newly introduced visual geometry benchmark (Tables 2–3; Section 6).

2. Context and Motivation¶

Problem / gap.
Many 3D vision tasks differ mainly by how many views you have (single image vs. multiple images/video) and whether camera poses are known, yet models are typically designed per-task (Introduction, paragraphs 1–2).
“Unified” models exist, but DA3 highlights several practical issues:
- They often use bespoke, complex architectures.
- They often require joint multi-task optimization from scratch, which makes it hard to fully leverage large-scale pretrained backbones (Introduction, paragraph 1).
Why it matters.
Recovering 3D structure supports robotics, mixed reality, and general spatial understanding (Introduction, paragraph 1).
A single model that works for Nv = 1 (monocular) up through many views would simplify deployment and evaluation across settings (Introduction; Figure 1).
Prior approaches and shortcomings (as positioned here).
Traditional SfM/MVS pipelines decompose the problem into matching → pose estimation → bundle adjustment → dense stereo, but can be brittle under low texture/specularities/large viewpoint changes (Section 2, “Multi-view visual geometry estimation”).
Transformer-based feed-forward geometry models (e.g., DUSt3R-style point maps; VGGT-style multi-output models) improve scalability and simplify pipelines, but can struggle with (i) arbitrary numbers of views and (ii) redundancy/entanglement in prediction targets (Section 2; Section 3.1 “Minimal prediction targets”).
How DA3 positions itself.
DA3 explicitly pursues minimal modeling: 1) A single plain transformer backbone (e.g., “vanilla DINOv2”) is sufficient (Introduction; Figure 2; Section 3.2). 2) A single depth–ray representation is a minimal but sufficient prediction target, reducing the need for multi-head, multi-task outputs (Section 3.1; Table 6).

3. Technical Approach¶

3.1 Reader orientation (approachable technical breakdown)¶

DA3 is a single transformer-based dense predictor that ingests one or many RGB images (and optionally camera poses) and outputs per-image depth maps and ray maps that are consistent across views (Figure 1; Section 3.1).
It solves “recover the visual space from any views” by turning multi-view geometry into a pixel-aligned prediction problem, then fusing predicted geometry into point clouds / reconstructions (Figure 1; Section 6.1).

3.2 Big-picture architecture (diagram in words)¶

Inputs: Nv images {I_i}; optionally camera parameters per view (intrinsics/extrinsics encoded as (t, q, f) in Section 3.1 / 3.2).
Backbone: a pretrained ViT (e.g., DINOv2) that processes image patch tokens (Section 3.2; Figure 2).
Cross-view reasoning: input-adaptive cross-view self-attention by rearranging tokens in selected layers so attention can mix information across views (Section 3.2).
Heads:
Dual-DPT head outputs per-view depth and per-view ray maps (Figure 3; Section 3.2).
Optional lightweight camera head predicts (t, q, f) from per-view camera tokens (Section 3.1, last paragraph; Section 3.2).
Outputs: depth maps + ray maps → 3D points via P = t + D * d (Section 3.1), enabling point clouds / 3DGS applications (Section 5).

3.3 Roadmap for the deep dive¶

I will explain: 1) The geometry formulation and why DA3 uses depth + ray instead of explicit rotation matrices (Section 3.1; Eq. (1)–(2); Table 6). 2) The model architecture, focusing on how a plain pretrained ViT is adapted to variable view counts (Section 3.2; Figure 2). 3) The prediction heads (Dual-DPT and optional camera components) and what they output (Figure 3; Section 3.1–3.2). 4) The training pipeline, including the teacher–student pseudo-label alignment and the loss terms (Section 3.3–3.4; Section 4; Eq. (3), (7), (8)). 5) The evaluation/benchmark pipeline and what the reported metrics actually measure (Section 6; Tables 2–5).

3.4 Detailed, sentence-based technical breakdown¶

Framing. This is primarily an empirical model + system design paper: it proposes a minimal transformer-based architecture plus a depth–ray prediction target, then demonstrates effectiveness via large-scale training, ablations, and new benchmarks (Introduction; Sections 3–7).

3.4.1 Formulation: from images to 3D via depth and rays¶

Inputs and standard projection.
The input is a set of images I = {I_i}_{i=1..Nv} with I_i ∈ R^{H×W×3} (Section 3.1).
Each view has depth D_i ∈ R^{H×W}, intrinsics K_i, and extrinsics (R_i | t_i) (Section 3.1).
A pixel p = (u, v, 1)^T can be mapped to a 3D point using depth and camera parameters (Section 3.1, first equations).
Why “depth–ray” instead of direct pose.
Predicting a valid rotation matrix R is hard because it must satisfy an orthogonality constraint (Section 3.1, “Depth-ray representation”).
DA3 avoids directly predicting R by predicting a per-pixel ray map that implicitly captures camera pose and projection.
Depth–ray representation (core idea).
For each pixel, DA3 defines a ray r ∈ R^6 as (t, d) where:
- t ∈ R^3 is the ray origin (camera center in world coordinates),
- d ∈ R^3 is the ray direction in world coordinates (Section 3.1).
The per-image ray map is M ∈ R^{H×W×6} storing (t, d) for all pixels (Section 3.1).
Importantly, DA3 does not normalize d, so its magnitude preserves projection scale (Section 3.1).
A 3D point is then computed by element-wise combination:
- P = t + D(u, v) · d (Section 3.1).
Micro-walkthrough (single pixel): 1) Take pixel coordinates p=(u,v,1). 2) The model predicts depth D(u,v) and ray direction d(u,v) and origin t(u,v) via M(u,v,:). 3) Compute point P(u,v) = t(u,v) + D(u,v) * d(u,v). 4) Repeating over pixels yields a point set that can be fused into a point cloud (Section 3.1; Section 6.1).
Recovering camera parameters from a ray map (optional inference step).
Camera center t_c is estimated by averaging ray origins across pixels (Eq. (1), Section 3.1).
Rotation and intrinsics are recovered by solving for a homography H = K R using a least-squares Direct Linear Transform objective:
- H* = argmin_{||H||=1} Σ || H p_{h,w} × M(h,w,3:) || (Eq. (2), Section 3.1),
- followed by RQ decomposition of H* to get K and R (Section 3.1).
The paper notes this ray→camera recovery is computationally costly at inference, motivating a lightweight explicit camera head (Section 3.1, “Minimal prediction targets”).
Minimal prediction targets claim (supported by ablations).
DA3 argues that depth + ray is a minimal sufficient set compared to alternatives like point maps or redundant multi-head targets (Section 3.1; Table 6 shows the comparative effect).

3.4.2 Architecture: a single pretrained transformer that scales with view count¶

Backbone: “single transformer” with no architectural specialization.
DA3 uses a pretrained Vision Transformer backbone (example given: DINOv2) as the feature extractor (Section 3.2; Figure 2).
The paper emphasizes there are no architectural changes to the transformer blocks themselves; cross-view behavior comes from token handling (Figure 2 caption; Section 3.2).
Handling arbitrary numbers of views via input-adaptive attention.
DA3 splits the L transformer blocks into:
- L_s layers of within-image self-attention, then
- L_g layers that alternate cross-view and within-view attention (Section 3.2).
The mechanism is implemented by rearranging tokens so the same self-attention operation can operate either per-view or across all views (Section 3.2).
The default configuration is L_s : L_g = 2 : 1 with L = L_s + L_g, chosen as the best performance/efficiency trade-off in ablations (Section 3.2; Table 7 discussion points to attention strategy differences, including “Full Alt.” vs partial).
Camera conditioning injection (optional).
Each view gets a camera token c_i prepended to its patch tokens (Section 3.2).
If camera parameters are available, a lightweight MLP E_c encodes (f_i, q_i, t_i) into c_i = E_c(f_i, q_i, t_i) (Section 3.2).
If camera parameters are not available, DA3 uses a shared learned token c_l as a placeholder (Section 3.2).
These camera tokens participate in attention, letting the model seamlessly handle posed/unposed settings (Figure 2; Section 3.2).

3.4.3 Output heads: Dual-DPT for depth+ray and optional camera head¶

Dual-DPT head (for pixel-aligned depth and rays).
DA3 predicts, for each input view, a depth map and a ray map aligned to the input image (Section 3.1; Figure 2).
The Dual-DPT head processes backbone features with:
- shared reassembly modules, then
- two separate sets of fusion layers, one for the depth branch and one for the ray branch,
- followed by separate output layers (Figure 3; Section 3.2).
The design goal is that both predictions use the same processed features but differ in final fusion, encouraging interaction without duplicating intermediate representations (Section 3.2; Figure 3 caption).
Optional camera head D_C.
DA3 can predict camera pose explicitly as v = (t, q, f) using a small transformer operating only on camera tokens (Section 3.1, last paragraph).
Because it processes one token per view, overhead is described as negligible; Table 6 discussion mentions ~0.1% of main backbone compute (Section 7.2.1 text following Table 6).

3.4.4 Training: teacher–student supervision + joint losses for depth, rays, points, and gradients¶

Why teacher–student is used.
Real-world depth sources can be noisy/incomplete (Figure 4; Section 3.3; Section 4.2).
DA3 trains a monocular “teacher” on synthetic data to generate dense pseudo-depth for real datasets, then aligns it to available (sparse/noisy) metric depth to preserve geometric correctness (Section 3.3; Section 4.2; Eq. (8)).
Training objective for DA3 (student any-view model).
The model F_θ maps inputs to {D̂, R̂, ĉ} (depth, ray map, optional camera output) (Section 3.3).
Before computing losses, all ground-truth signals are normalized by a common scale: the mean ℓ2 norm of valid reprojected point maps P (Section 3.3). (The paper motivates this as stabilizing magnitudes across modalities.)
The overall loss is a weighted sum (Section 3.3):
- L = L_D(D̂, D) + L_M(R̂, M) + L_P(D̂ ⊙ d + t, P) + β L_C(ĉ, v) + α L_grad(D̂, D)
- with weights α = 1 and β = 1 (Section 3.3).
The gradient loss is explicitly (Eq. (3), Section 3.3):
- L_grad = ||∇x D̂ − ∇x D||_1 + ||∇y D̂ − ∇y D||_1,
- intended to preserve sharp edges while keeping planar regions smooth (Section 3.3).
Depth loss L_D uses an ℓ1 term modulated by per-pixel confidence D_c,p and includes a -λ_c log D_c,p regularizer (Section 3.3). (The value of λ_c is not provided in the excerpted text.)
Training schedule and compute (DA3).
Hardware/steps: trained on 128 H100 GPUs for 200k steps (Section 3.4).
LR schedule: 8k-step warm-up, peak LR 2 × 10^-4 (Section 3.4).
Resolution strategy:
- Base resolution 504×504 (chosen to be divisible by many factors for aspect ratios) (Section 3.4),
- Training samples multiple resolutions: 504×504, 504×378, 504×336, 504×280, 336×504, 896×504, 756×504, 672×504 (Section 3.4).
View count: for 504×504, Nv is sampled uniformly from [2, 18] (Section 3.4).
Batch sizing: dynamically adjusted to keep token count per step roughly constant (Section 3.4).
Supervision transition: switches from ground-truth depth to teacher labels at 120k steps (Section 3.4).
Pose conditioning is randomly activated with probability 0.2 during training (Section 3.4).
Optimizer / ViT internal hyperparameters.
- The excerpt does not specify DA3’s optimizer type (e.g., AdamW) or standard ViT configuration details (layers, heads, hidden size) for the any-view model; it reports model scales via parameter counts in Tables 2–3 and speed in Table 8.

3.4.5 Teacher model construction and pseudo-label alignment¶

Teacher architecture and target.
The teacher is a monocular relative depth model using a DINOv2 ViT backbone plus a DPT decoder, explicitly described as aligned with DA3’s “plain transformer + DPT” philosophy (Section 4.1).
The teacher predicts scale–shift-invariant depth (not disparity), and predicts exponential depth to increase near-range discrimination (Section 4.1, “Depth representation”).
Teacher losses (geometric detail).
Teacher objective combines:
- gradient loss,
- a global–local loss (L_gl) referenced from prior work,
- a distance-weighted surface-normal loss defined via Eqs. (4)–(6),
- plus sky-mask and object-mask losses (MSE) (Section 4.1; Eq. (7)).
The combined teacher loss is (Eq. (7), Section 4.1):
- L_T = α L_grad + L_gl + L_N + L_sky + L_obj with α = 0.5.
Data scaling for the teacher.
The teacher is trained exclusively on synthetic data, and the paper lists an expanded synthetic corpus spanning indoor/outdoor/object-centric/in-the-wild-like scenes (Section 4.1, “Data scaling”; also Table 1 lists many training datasets).
Aligning teacher pseudo-depth to real metric depth.
Let D~ be teacher relative depth and D be sparse/noisy depth with mask m_p.
DA3 fits scale s and shift t via a RANSAC least-squares procedure (Eq. (8), Section 4.2):
- (ŝ, t̂) = argmin_{s>0,t} Σ m_p (s D~_p + t − D_p)^2,
- then uses D_{T→M} = ŝ D~ + t̂.
The inlier threshold is defined as the mean absolute deviation from the residual median (Section 4.2).

4. Key Insights and Innovations¶

(1) Depth + ray map as a minimal sufficient prediction target for any-view geometry.
Novelty relative to point-map-only or redundant multi-output approaches: DA3 encodes pose implicitly via dense rays, avoiding direct rotation matrix constraints while still enabling consistent 3D point computation P = t + D·d (Section 3.1).
Significance: Ablations show depth + ray outperforms depth + cam and depth + pcd + cam across datasets (Table 6; Section 7.2.1), suggesting rays provide a strong “single representation” for both structure and camera motion.
(2) A single plain pretrained transformer backbone is sufficient (no bespoke multi-stage architecture required).
DA3 emphasizes “vanilla DINOv2” and token-rearrangement-based cross-view reasoning, rather than stacking multiple transformers or introducing heavy specialized modules (Figure 2; Section 3.2).
The ablation comparing the proposed single-transformer design vs. a “VGGT style” stacked-transformer design reports large degradation for the stacked variant in their setting (Table 7, rows a vs b; Section 7.2.2).
(3) Input-adaptive cross-view attention via token rearrangement supports arbitrary view counts.
The mechanism is conceptually simple: choose which layers do within-view vs cross-view attention by changing how tokens are grouped (Section 3.2).
Significance: It allows DA3 to naturally reduce to monocular depth estimation when Nv=1 without extra overhead (Section 3.2).
(4) Dual-DPT head for aligned depth and ray prediction.
The “shared reassembly + branch-specific fusion” design is intended to improve alignment/interaction between depth and ray predictions without duplicating full decoders (Figure 3; Section 3.2).
Ablation shows performance drops when replacing Dual-DPT with two separate DPT heads (Table 7, row d; Section 7.2.3).
(5) Teacher-only-on-synthetic + robust scale/shift alignment as a practical data unification strategy.
The teacher produces dense relative depth even where real depth is sparse/noisy (Figure 4; Section 3.3).
The robust alignment (Eq. (8)) is the mechanism that attempts to preserve metric consistency while improving detail (Section 4.2), and the qualitative example suggests improved fine structures with teacher supervision (Figure 8).

5. Experimental Analysis¶

5.1 Evaluation methodology (datasets, metrics, setup)¶

Visual Geometry Benchmark (new).
Datasets: HiRoom (synthetic), ETH3D, DTU, 7Scenes, ScanNet++ (Section 6.3; Table 1 lists counts).
Pose evaluation:
- Uses AUC over thresholds derived from Relative Rotation Accuracy (RRA) and Relative Translation Accuracy (RTA) (Section 6.2).
- Reports AUC at thresholds 3 and 30 (Section 6.2; Table 2 columns Auc3 and Auc30).
Geometry/reconstruction evaluation:
- Reconstruct point clouds by fusing predicted poses + predicted depths, aligning to ground truth via evo alignment plus a RANSAC robustness scheme (Section 6.1).
- TSDF fusion is used for reconstruction (Section 6.1).
- Metrics: F1-score at a dataset-specific distance threshold d; DTU reports Chamfer Distance (mm) (Section 6.2; Table 3).
Feed-forward novel view synthesis benchmark (new).
Uses DL3DV (140 scenes), Tanks and Temples (6), MegaDepth (19) (Section 6.3; Table 5).
Protocol: 12 context views selected; evaluate on target views sampled every 8; image resolution 270×480 (Section 6.1; Table 5 caption).
Metrics: PSNR / SSIM / LPIPS (Section 6.3; Table 5).

5.2 Main quantitative results¶

Pose accuracy (Table 2).
DA3-Giant achieves best reported AUCs across all listed datasets for Auc3 and for most Auc30 entries.
Concrete examples from Table 2:
- HiRoom: DA3-Giant Auc3=80.3, Auc30=95.9 vs VGGT 49.1/88.0 and Pi3 67.0/94.8.
- ETH3D: DA3-Giant 48.4/91.2 vs VGGT 26.3/80.8 and Pi3 35.2/87.3.
- ScanNet++: DA3-Giant 85.0/98.1 vs VGGT 62.6/95.1 and Pi3 50.7/92.1.
The text notes the only exception is Auc30 on DTU (Section 7.1); Table 2 shows VGGT has 99.8 on DTU Auc30 while DA3-Giant has 99.4.
Reconstruction / geometry accuracy (Table 3).
Pose-free (w/o p.) performance: DA3-Giant is best across HiRoom, ETH3D, DTU (CD), 7Scenes, ScanNet++ in the table.
- HiRoom w/o pose: DA3-Giant F1=85.1 vs VGGT 56.7, Pi3 75.8.
- ETH3D w/o pose: DA3-Giant F1=79.0 vs VGGT 57.2, Pi3 72.7.
- DTU w/o pose (CD↓): DA3-Giant 1.85 mm vs VGGT 2.05 mm, Pi3 3.28 mm.
With ground-truth pose (w/ p.), DA3-Giant remains very strong, e.g. HiRoom F1=95.6, ETH3D F1=87.1, ScanNet++ F1=79.3 (Table 3).
Monocular depth (Table 4).
DA3 reports δ1 on KITTI/NYU/SINTEL/ETH3D/DIODE and a rank summary.
From Table 4:
- KITTI δ1: DA3 95.3 vs DA2 94.6.
- ETH3D δ1: DA3 98.6 vs DA2 86.5 (large gap in the table).
- Teacher is strongest overall in that table (e.g., KITTI 97.2, ETH3D 99.8) (Table 4).
Feed-forward NVS with 3DGS (Table 5).
DA3 as backbone (“DAv3 (Ours)”) achieves the best numbers across all three datasets in Table 5:
- DL3DV: PSNR=21.33, SSIM=0.711, LPIPS=0.241 (best among listed).
- Tanks and Temples: 18.10 / 0.578 / 0.311.
- MegaDepth: 17.89 / 0.561 / 0.351.
The text emphasizes out-of-domain drops (DL3DV vs Tanks/MegaDepth) as a sensitivity to trajectory/pose distributions (Section 7.1).

5.3 Do the experiments support the claims?¶

Support for “depth+ray is sufficient / minimal.”
Table 6 directly tests output-target combinations and shows depth + ray substantially improves pose (Auc3) and geometry metrics over depth + cam and depth + pcd + cam across datasets (Section 7.2.1; Table 6).
Adding an auxiliary camera head (depth + ray + cam) does not consistently improve over depth + ray in Table 6, which supports the “ray sufficiency” argument (Section 7.2.1 text).
Support for “single transformer suffices.”
Table 7 compares the proposed architecture to a “VGGT style” stacked-transformer approach under similar parameter sizing (Section 7.2.2). The stacked variant is dramatically worse in that ablation table (Table 7 row b vs row a).
Support for teacher-student value.
Table 7 row (e) “w/o Teacher” shows mixed changes (e.g., improved DTU CD but worse elsewhere), while Figure 8 qualitatively shows more detailed depth with teacher supervision (Section 7.2.3; Figure 8).

5.4 Ablations, failure cases, robustness checks¶

Ablations included:
Output target combinations (Table 6).
Attention scheduling (“Full Alt.” vs partial) and architecture variants (Table 7 a–c).
Dual-DPT head vs separate heads (Table 7 d).
Teacher supervision on/off (Table 7 e; Figure 8).
Pose conditioning on/off (Table 7 f–g; note these rows are evaluated with ground-truth pose fusion, marked * in Table 7).
Robustness / scaling:
Table 2 and Table 3 report multiple DA3 sizes (Small/Base/Large/Giant) with parameter counts, giving a partial view of scaling trends.
Table 8 reports max number of images and FPS at 504×336 (Table 8).
Failure cases:
The provided excerpt does not include explicit failure-case categories; it mainly provides positive qualitative comparisons (Figures 6–7, 9–11).

6. Limitations and Trade-offs¶

Compute and accessibility.
Training DA3-Giant is expensive: ~10 days on 128×H100 GPUs (Section 7.2).
Even ablations are reported as ~4 days on 32×H100 with ViT-L and max 10 views (Section 7.2), which limits reproducibility for many groups.
Ray-map-to-camera recovery cost vs convenience.
Recovering (K,R,t) from rays involves solving a DLT least-squares problem (Eq. (2)) and RQ decomposition (Section 3.1), and the paper explicitly says this is “computationally costly” at inference (Section 3.1).
DA3 therefore adds a camera head for practical use (Section 3.1), but that means the system is no longer purely “depth+ray only” in deployed form (even if the camera head is lightweight).
Dependence on teacher pseudo-labeling and alignment assumptions.
The approach assumes teacher-relative depth can be robustly aligned to noisy metric depths with a global scale+shift per image (Eq. (8), Section 4.2). If real depth is extremely sparse/biased, this alignment could be unreliable (the paper does not enumerate such cases in the excerpt).
Teacher is trained exclusively on synthetic data (Section 4.1), which is a deliberate choice to avoid poor real depth (Figure 4) but could introduce synthetic-to-real generalization gaps; DA3 mitigates this by aligning to real measurements (Section 4.2), though the residual domain gap is not fully characterized here.
Metric depth vs relative depth scope.
The main DA3 any-view formulation predicts depth and rays with normalization steps and optional pose conditioning; metric depth is treated as a separate model variant (“DA3-metric”) with its own training protocol and canonical focal length transformation (Section 4.4; Section 7.4).
Benchmark pipeline choices may affect conclusions.
Reconstruction uses evo-based pose alignment plus RANSAC selection, then TSDF fusion (Section 6.1). This couples final geometry scores to (i) alignment heuristics and (ii) fusion settings (e.g., dataset-specific TSDF voxel sizes in Section 6.3).
DTU evaluation uses background removal (RMBG 2.0) and a specific fusion strategy (Section 6.3), which may advantage methods that behave well under those preprocessing steps.
Missing low-level implementation specifics (from the provided excerpt).
For DA3 training, the optimizer type, weight decay, exact batch sizes per resolution/view-count, and full ViT block hyperparameters (layers/heads/hidden size per variant) are not specified in the shown sections (Section 3.4 provides steps/LR/hardware/resolutions but not all optimizer details).

7. Implications and Future Directions¶

How this changes the field (as evidenced in this paper).
DA3 suggests that unified visual geometry can be effectively approached as dense per-view prediction (depth + rays) using a single pretrained ViT, rather than designing separate SfM/MVS/monocular architectures (Introduction; Section 3; Tables 2–3).
The new benchmark explicitly scores both pose and reconstructed geometry, encouraging holistic evaluation rather than task-isolated metrics (Section 6; Tables 2–3).
Follow-up research directions suggested by the work.
The conclusion points to extending to dynamic scenes, adding language/interaction cues, and scaling pretraining further toward actionable world models (Section 8).
Practical applications / downstream use cases demonstrated here.
Feed-forward novel view synthesis (FF-NVS): DA3 is fine-tuned with an added DPT head to predict pixel-aligned 3D Gaussian parameters for 3DGS rendering (Section 5.1), and a pose-adaptive variant supports posed or unposed inputs (Section 5.2). Quantitative gains on the NVS benchmark suggest geometry strength transfers to rendering quality (Table 5; Section 7.1).
Monocular depth estimation: even though DA3 is designed for any-view geometry, it reports strong δ1 results on monocular benchmarks (Table 4).
Repro/Integration Guidance (grounded in what the paper describes).
Prefer DA3’s pose-free mode when camera poses are unavailable; DA3 still predicts rays and (optionally) camera pose via its camera head (Sections 3.1–3.2; Tables 2–3 w/o p.).
Prefer pose conditioning when accurate camera parameters are available: pose conditioning improves reconstruction metrics in the ablation that compares “w/ Pose Cond.” vs “w/o Pose Cond.” (Table 7 f–g, noting these are evaluated with ground-truth pose fusion).
For NVS/3DGS pipelines, the paper’s minimal recipe is: freeze the DA3 backbone and train only the added GS-DPT head for stability (Section 5.2), and train with varying resolutions/view counts (Section 5.2).