
Depth Pro 2 vs VGGT: Metric Depth or Multi-View Geometry
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 186 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 358 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 87 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 139 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-0737 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 90 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 135 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 80 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 117 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 111 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 66 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 91 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 42 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 210 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 109 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens · 168 tok/s
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
Point a phone at a room, walk around it, and record forty seconds of video. Two models can turn that into geometry, and they will not agree on what the geometry is. Depth Pro 2, Apple's metric monocular depth model, returns a depth value in meters for every pixel of every frame, with a focal length attached. VGGT — the Visual Geometry Grounded Transformer from Meta AI Research and the University of Oxford's VGG group, published at CVPR 2025 — returns camera extrinsics and intrinsics, point maps, depth maps and 3D point tracks for the whole set of views at once.
One of those outputs is a measurement. The other is a reconstruction. Choosing between them is not a question of which is more accurate, because the two models are answering different questions, and the answer you need is decided by what comes after the model in your pipeline.
What each model was built to do
VGGT is a feed-forward network that infers all the key 3D attributes of a scene from one, a few, or hundreds of views, within seconds. The pitch is joint estimation: rather than running a depth model per frame and then solving for camera poses with a separate structure-from-motion stage, VGGT produces poses, intrinsics, depth and point tracks in a single pass, with the views attending to each other. That is why the "hundreds of views" claim matters — it is a replacement for a photogrammetry pipeline, not an addition to it.
Depth Pro 2 is a monocular metric depth estimator, and the word doing the work is metric. Its output is a depth map in meters along the camera's viewing direction, computed from a single image, with no camera metadata required — the model can read a focal length from EXIF, accept one you supply in pixels, or predict it with a separate 23.5M-parameter estimator that ships as its own repository. It also handles video, which is where the "2" comes from: a multi-view transformer processes frames jointly, alternating attention within each frame and across frames after the first third of its blocks, and a DPT decoder emits an inverse depth map per frame.

• Unit of work — VGGT takes a set of views; Depth Pro 2 takes a photo, a video, or a live feed.
• Primary output — VGGT gives camera parameters, point maps, depth maps and 3D point tracks; Depth Pro 2 gives metric depth plus the focal length it used.
• Camera solve — VGGT estimates it as part of the joint problem; Depth Pro 2 reads, accepts or predicts a focal length, without extrinsics.
• Scale — Depth Pro 2's headline output is meters; VGGT's model card describes camera parameters, point maps, depth and tracks and does not claim metric scale.
• Release — VGGT was published at CVPR 2025 (arXiv 2503.11651); Depth Pro 2's repositories were created in November 2026, with a technical report still pending.
The scale question decides most deployments
If you need to know that the table is 0.74 meters away, VGGT does not directly tell you that, and Depth Pro 2 does. That is the whole argument in one sentence, and it is worth being precise about why.
Depth Pro 2's metric claim is architectural, not incidental. The original Depth Pro was built as a foundation model for zero-shot metric depth, and the repository is explicit about the conventions: depth is the z coordinate along the viewing direction rather than radial distance, inverse depth is normalized by focal length, and a focal length 10 percent too long places everything 10 percent too far. Those are the conventions of a model whose output is meant to be read as a measurement, and Depth Pro 2 inherits them unchanged across all four sizes.
VGGT's output is geometric. Camera extrinsics, intrinsics, point maps and point tracks are internally consistent — the reconstruction is correct up to a global similarity transform — and recovering absolute scale from it is a separate step, usually involving a known reference in the scene. For novel-view synthesis, for a 3D asset, for anything where you will rescale the scene to fit a viewer, that is fine and often preferable. For a robot that needs to know whether an obstacle is one meter or three meters away, it is a problem you have to solve outside the model.
Depth Pro 2 has its own honesty problems here. Sky pixels come back as infinite depth, or, especially on the smaller models, as a finite depth of several kilometers, so a naive point cloud built from its output will have a wall of spurious far geometry where the sky is. The README tells you to clip the depth range before making a point cloud or a color image, which is the kind of instruction that only appears after someone has been bitten by it.
Video consistency is the real contribution
Depth Pro 2 is not a better Depth Pro so much as a temporally stabilized one, and the mechanism is entirely in the attention pattern. After the first third of its blocks, the transformer alternates attention within each frame and attention across frames, and the DPT decoder assembles tokens from several blocks into feature maps at one thirty-second to one quarter of the input resolution. The point of that alternation is Jitter — the temporal-consistency metric — and the project page reports it alongside Boundary F1, averaged over the Sintel and Spring datasets, with the claim that Depth Pro 2 outperforms prior work operating on a single frame or multiple frames on both.
No values accompany that claim, and no baselines are named. The page refers only to "prior work operating on a single frame or multiple frames" and to "baselines at their default settings," and the citation block reads arXiv:261X.XXXXX — a placeholder, not a paper. Anyone quoting a Boundary F1 or Jitter number for Depth Pro 2 is reading it off a chart.

Depth Pro 2's operating envelope is narrower than its video framing suggests, and the narrowness is documented. Its default mode consumes 16-frame clips with four frames of overlap, and the live path runs each frame against the fifteen that preceded it, which means the first second of any stream is the shakiest second you will get. One focal length serves an entire video, picked as the median over the opening window — so a clip that zooms will have one frame's worth of focal estimate applied to all of them. Memory scales with frame size and window, not with duration, and Apple's remedy for an out-of-memory video is to shorten the window before shrinking the model.
VGGT has no equivalent mode, because it does not need one. It already handles the whole view set jointly, which is a stronger form of the same guarantee — provided you can afford the attention over hundreds of views, and provided your application wants a reconstruction rather than a per-frame depth reading.
Sizes, exports and what you actually run
Depth Pro 2 gives you four sizes with published counts: S at 24M parameters and 97 MB, B at 97M and 386 MB, L at 334M and 1.3 GB, and H+ at 872M and 3.5 GB, with H+ as the default and S recommended for laptops. The configuration files show the family structure — DINOv3 ViT-S and ViT-B backbones on the small pair attending at blocks 5, 7, 9 and 11, ViT-L attending at blocks 9 through 23, ViT-H+ at blocks 11 through 31, with decoder dimensions of 64, 128, 256 and 256. Core ML exports for Apple silicon ship as a separate repository, and the code repository includes a Mac demo app for live depth from the camera.
VGGT-1B is a single billion-parameter checkpoint, and the family has since moved on. VGGT-Omega was released on 18 May 2026 as "the next step of VGGT," and a commercial checkpoint, VGGT-1B-Commercial, exists alongside the original with an AUC@30 of 90.37 on Co3D against 89.98 for the original — the only benchmark number on the model card, and a marginal one. VGGT-1B carries 748,148 downloads as of this writing, which tells you something about how widely the research community adopted it as a baseline.

Depth Pro 2 is the opposite story. Its four checkpoints have downloads in the tens, the repositories were created in early November 2026, and the only likes on the card repository are a handful. It is a model that has been published and not yet picked up. That is worth knowing before you commit: if you hit a bug, you will be the one filing the issue.
Licensing and the commercial checkpoint
VGGT-1B is CC BY-NC 4.0, non-commercial. The existence of VGGT-1B-Commercial is the interesting detail — Meta shipped a commercially licensable variant, which means the path from research to product exists without a license negotiation. Depth Pro 2's weights are under the Apple Machine Learning Research Model license, research use only, defined by the repository as non-commercial scientific research and academic development activities.
Neither license is a product license on the open checkpoint. The difference is that VGGT's family includes a commercial option and Depth Pro 2's does not, so if you are building for revenue, one of these has a door and the other has a wall.
This is the same calculation that shows up one layer up in any production pipeline. A reconstruction model is one call; around it sits the part that reads a scene, decides what to do, and writes it down — language-model work, where the vendor lists a price and an aggregator either passes it through or does not. OrcaRouter's pitch is 0% markup on 217 models with automatic failover across providers, which matters most exactly here: when the geometry stage is a research checkpoint with a handful of downloads and no release cadence, you want the rest of the pipeline to be the boring, redundant part.
Who should start with which
Start with VGGT if you are reconstructing a scene from many views and you will handle scale downstream — novel-view synthesis, a 3D asset, a dataset built from video where the geometry is the deliverable. The joint camera-and-geometry solve is what makes it worth the billion parameters, and VGGT-Omega is the current generation if you are starting fresh.
Start with Depth Pro 2 if you need meters from a single frame, or if you need depth to stay steady frame to frame in a live feed. The size dial is real and the laptop workflow is documented. Go in expecting to clip sky depth, expect to be the first person to file most bugs, and expect to wait for the technical report before you can cite a Boundary F1 number to anyone.
What you should not do is treat this as a head-to-head. VGGT reconstructs, Depth Pro 2 measures, and the only honest comparison is which one your downstream code can consume. For most 3D pipelines the answer is VGGT and a scale recovery step; for most robotics and AR pipelines it is Depth Pro 2 and a robustness budget for the smoke.
