
Depth Pro 2 vs MoGe-2: Video Consistency vs Point Maps
- 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
Depth Pro 2 and MoGe-2 are both metric monocular geometry models from large research labs, and as of this writing nobody has published a number that puts one against the other. Apple's Depth Pro 2 reports Boundary F1 and Jitter, averaged over the Sintel and Spring datasets. Microsoft's MoGe-2 reports relative geometry accuracy, metric scale, and detail preservation on its own evaluation mix. The two sets of metrics do not intersect, which means every "X beats Y" claim you will find about this pairing is either invented or borrowed from a benchmark neither lab ran.
That is not a dodge. It is the single most useful thing to know before you spend a week integrating either one. What you can compare is what each repository actually ships, what each model was designed to be bad at, and which of those weaknesses lands on your pipeline. This piece does that, and it is explicit about which claims come from a vendor README and which come from nothing at all.
What each repository actually contains
Depth Pro 2 lives across several Hugging Face repositories rather than one. The card repository, apple/DepthPro2, holds the README and the figures; the weights are split by size into apple/DepthPro2-S, apple/DepthPro2-B, apple/DepthPro2-L and apple/DepthPro2-H. A separate repository, apple/DepthPro2-focal, carries a standalone 23.5M-parameter focal-length estimator that loads alongside any size. Another, apple/DepthPro2-coreml, holds Core ML exports for Apple silicon. The code repository is apple-aiml-research/ml-depth-pro-2 on GitHub.

The four sizes run from 24M parameters at 97 MB (S) through 97M at 386 MB (B) and 334M at 1.3 GB (L) to 872M at 3.5 GB (H+, the default). Their configurations confirm the family relationship: S and B use DINOv3 ViT-S and ViT-B backbones with cross-frame attention at blocks 5, 7, 9 and 11; L widens that to blocks 9 through 23 with a 256-dimension DPT decoder; H+ goes to blocks 11 through 31 and assembles tokens at blocks 7, 15, 23 and 31.
MoGe-2 ships as four checkpoints under the Ruicheng organization: moge-2-vitl at 326M parameters without normals, and the -normal variants moge-2-vitl-normal at 331M, moge-2-vitb-normal at 104M, and moge-2-vits-normal at 35M. Each returns a metric point map in OpenCV camera coordinates, a depth map, normalized camera intrinsics, and a valid-pixel mask, with the normal map added on the variants that carry it. MoGe-2 is the second generation; the family has since moved on to MoGe-3, announced 18 August 2026 with a 1.25B-parameter ViT-g model, so MoGe-2 is now the previous release rather than the current one.

Where the two designs diverge
• Input modality — Depth Pro 2 takes one photo, a video file, or a live camera feed; MoGe-2 takes a single image.
• Headline claim — Depth Pro 2 claims sharpness and temporal consistency; MoGe-2 claims metric scale plus fine detail, achieved through a unified data refinement step that uses sharp synthetic labels to filter and complete real training data.
• Output surface — Depth Pro 2 returns depth in meters plus the focal length it used; MoGe-2 returns a full point map, depth, intrinsics and a mask, with normals on the variants that include them.
• Focal length — Depth Pro 2 ships a separate estimator and can also read EXIF or accept a known focal length in pixels; MoGe-2 can take field of view as optional ground truth to improve accuracy.
• Scale — MoGe-2 added real-world metric scale over MoGe-1, which produced affine-invariant point maps with unknown scale; Depth Pro 2 inherits metric scale from the original Depth Pro.
• License — Depth Pro 2 weights are under the Apple Machine Learning Research Model license, research use only, defined as non-commercial scientific research and academic development. MoGe-2 code is MIT, with the bundled DINOv2 under Apache 2.0.
The video question is the whole argument
Depth Pro 2's distinguishing architectural decision is that a multi-view transformer processes one or more frames jointly, and after the first third of its blocks it alternates attention within each frame and attention across frames. A DPT decoder then predicts an inverse depth map per frame. That is the machinery behind the Jitter number — the temporal-consistency metric — and it is why the model has a video story at all.
In practice that means two operating modes. The default, buffered mode, processes clips of 16 frames that overlap by four, which keeps depth steady where clips meet; the window and overlap settings expose the clip length and the overlap. Unbuffered mode predicts each frame as it arrives using that frame and the 15 before it, which is the mode you want for a camera. Apple's own documentation is candid about the cost of the live path: the first frames of a stream have fewer frames before them to attend to, so their depth is less steady. A video gets one focal length for all its frames, taken as the median of the estimates over its first window.
MoGe-2 has no equivalent. It is a single-image model, and running it on video means running it frame by frame and accepting whatever flicker falls out. For a static reconstruction task — photogrammetry, a still-to-3D asset pipeline, anything where each frame is independent — that is not a limitation at all. For a robot or an AR overlay, it is the difference between usable and not.
What "metric" means in each output
Both models claim meters, and both are doing something slightly different with the word. Depth Pro 2 returns distance along the camera's viewing direction — the z coordinate — not the straight-line distance from the camera. It also normalizes inverse depth by focal length, so inverse depth equals the focal length in pixels divided by the product of image width and depth. Sky pixels come back as infinite depth, or, especially on the smaller models, as a finite depth of several kilometers, which is a detail worth handling before you build a point cloud.
MoGe-2 returns a point map directly, which sidesteps that conversion. That is the more useful output if you are going to 3D anyway, and it is why the model reports intrinsics alongside depth — the point map is only meaningful in a camera frame you also know. Neither output is a substitute for a real depth sensor when the scene has no texture to anchor scale; both are estimating, and both will drift on reflective or transparent surfaces.
Licensing is the decision, not the benchmark
Neither model is commercially usable as shipped, but the constraints differ in ways that matter to a legal review. Depth Pro 2's weights are under the Apple Machine Learning Research Model license, which permits use exclusively for research purposes, with the repository defining that as non-commercial scientific research and academic development activities. The code is separately licensed, and the repository acknowledges that it builds on TIMM and DINOv3.
MoGe-2's code is MIT, which is close to unencumbered, though the weights are not given separate terms on the model card and the bundled DINOv2 code is Apache 2.0. If your question is "can I ship this in a product," the honest answer for both is that you need a conversation with legal, and the MoGe-2 conversation is shorter. If your question is "can I publish a paper with this," both are fine.

What the missing numbers should tell you
Depth Pro 2's technical report does not exist yet. The repository's citation block still reads arXiv:261X.XXXXX, and the Xs are not a redaction — they are a placeholder that has not been filled in. The project page names the metrics it claims to win, Boundary F1 and Jitter, and names no baselines, referring only to "prior work operating on a single frame or multiple frames" and to "baselines at their default settings."
MoGe-2 has a real paper, arXiv 2507.02546, but its abstract is equally number-free; evaluation is deferred to a separate document. So the comparison table you were probably hoping for does not exist on either side, and any article that hands you one is quoting a chart reading.
Which is worth keeping in mind when you price the rest of the stack. A depth model is one stage of a pipeline that also has to describe the scene, decide what to do about it, and write the result somewhere — and those stages are language-model calls, where pricing is published and directly comparable. That is the argument for putting them behind one API rather than three: OrcaRouter routes 217 models on a single key at 0% markup, meaning the provider's list price passes through untouched, so a vendor price cut is live the same day rather than whenever an aggregator re-syncs. Automatic failover across providers matters for the same reason in reverse — if the reasoning stage is the one that has to stay up, you want the depth job to be the only thing in the pipeline allowed to fail quietly.
Which one to reach for
If you have video and you care about flicker, Depth Pro 2 is the only one of these two with an answer, and its answer is real architecture — cross-frame attention, overlapping windows, a live mode with a documented warm-up penalty. Start with S on a laptop, as Apple's own documentation suggests, and move up only when the edges are not sharp enough.
If you have stills and you want a point map, MoGe-2 gives you more output surface per call, and the normal variants give you a second signal you would otherwise estimate. Its metric scale is the improvement over MoGe-1, and it holds.
What neither gives you is a defensible claim that it is better than the other, because neither lab ran the experiment. Treat the pairing as two different tools with overlapping names, not as a ranked list, until somebody publishes the numbers.
