
Depth Pro 2 vs UniDepth V2: One Model or a Whole Family
- 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
The interesting thing about Depth Pro 2 is not that it estimates depth. UniDepth V2 does that, reports it on ten datasets in a zero-shot regime, and has done so since its paper appeared as arXiv 2502.20110 in February 2025. The interesting thing is that Depth Pro 2 refuses to be one model. Apple shipped it as four sizes plus a focal-length estimator plus a Core ML export, and the size you pick changes what the model can do to you almost as much as which model you picked.
Both are metric monocular estimators trained to work on images they have never seen. Depth Pro 2 returns depth in meters along the camera's viewing axis; UniDepth V2 returns metric depth and a point cloud in camera coordinates, along with the intrinsics it inferred. Both estimate the camera rather than requiring it. The comparison that matters is not accuracy — the two labs have not run a shared benchmark — but how much of the deployment decision each one leaves to you.
Four sizes against three backbones
UniDepth V2's model zoo is three checkpoints: unidepth-v2-vits14, unidepth-v2-vitb14 and unidepth-v2-vitl14, one per ViT backbone size. The paper describes a self-promptable camera module that predicts a dense camera representation to condition the depth features, a pseudo-spherical output that separates camera and depth representations, an edge-guided loss inherited from UniDepth V1, and an uncertainty output. The model accepts known intrinsics or a camera class — Pinhole, Fisheye624 — and skips the prediction when you supply them.
Depth Pro 2's four sizes are 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. The configuration files make the scaling concrete: S and B run DINOv3 ViT-S and ViT-B backbones and attend across frames at blocks 5, 7, 9 and 11; L runs ViT-L and attends at blocks 9 through 23; H+ runs ViT-H+ and attends at blocks 11 through 31. The decoder dimension doubles from 64 on S to 128 on B and 256 on both L and H+.

• Parameter spread — UniDepth V2 spans three ViT sizes with no published counts on the model card; Depth Pro 2 spans 24M to 872M with the counts in the README.
• Default checkpoint — Depth Pro 2 defaults to H+ at 3.5 GB; UniDepth V2 leaves the choice to the caller.
• Export targets — Depth Pro 2 ships Core ML packages for Apple hardware and a separate 23.5M-parameter focal-length estimator; UniDepth V2 ships ONNX support.
• Camera model — UniDepth V2 exposes Pinhole and Fisheye624 classes and accepts known intrinsics; Depth Pro 2 assumes a perspective camera, reads EXIF, or estimates focal length with the separate model.
• Uncertainty — UniDepth V2 returns a confidence output; Depth Pro 2 does not.

The metric-scale lineage runs the other way
UniDepth V2 is a refinement of a metric-depth design, and the paper is explicit that V2's job was to make the architecture simpler and more efficient while adding sharper edges through the edge-guided local SSI, flexible input shapes and ratios via a resolution level setting, faster inference, ONNX support, and the new camera models. The reported zero-shot numbers, where higher is better on the delta-1 metric, are 98.8 on NYUv2, 96.4 on SUN-RGBD, 85.2 on ETH3D, 94.5 on iBims-1, 98.9 on KITTI, 87.0 on Nuscenes and 88.2 on DDAD, across ten datasets in total. Those are vendor-reported and unreproduced here.
Depth Pro 2 comes at metric scale from the other direction. The original Depth Pro was a foundation model for zero-shot metric depth that produced 2.25-megapixel maps in 0.3 seconds and estimated focal length from a single image, and Depth Pro 2's contribution is not scale but consistency — sharp boundaries and steady depth across frames. There is no published delta-1 table for it. The project page names Boundary F1 and Jitter as the metrics it claims to win and gives no values, and the citation block still points at arXiv:261X.XXXXX, a placeholder rather than a paper.
So the two models are not competing on the same axis. UniDepth V2 is a depth estimator with a strong, tabulated zero-shot story. Depth Pro 2 is a depth estimator that has been made temporally stable, and the evidence for that is a chart and a claim. If your evaluation is "which has numbers I can put in a procurement document," UniDepth V2 has them and Depth Pro 2 does not yet.
What the output actually gives you
The difference in output shape is larger than it looks. UniDepth V2 hands you a point cloud already expressed in camera coordinates, with the intrinsics that make it meaningful and an uncertainty channel you can threshold on. Depth Pro 2 hands you a depth map plus the focal length it settled on, which means the conversion to 3D is yours to write and yours to get wrong. If you are producing assets, UniDepth V2 saves a stage. If you are producing depth — a matte, an AR occlusion mask, a focus-pull — the extra structure is work you did not need.
Neither model replaces a depth sensor on a textureless or reflective surface. Both are estimating scale from visual evidence, and both will be wrong in the same places for the same reasons.
Licensing, and the commercial question
UniDepth V2 is released under Creative Commons BY-NC 4.0, which is non-commercial. Depth Pro 2's weights are under the Apple Machine Learning Research Model license, which the repository defines as research use only — non-commercial scientific research and academic development. On the licensing question the two are in the same place: neither is a product license, and both need a legal conversation before they touch revenue.
The difference is what surrounds them. UniDepth V2's code repository is the reference implementation and the model card points at it for documentation; Depth Pro 2 ships a pip-installable package with a command line, a Python API, a Mac demo app, and an explicit statement that the repository contains code to run the models but not training or evaluation code. If you are evaluating "how fast can a team be productive with this," Depth Pro 2 has done more of that work.

Where this lands in a real system is usually the same place every time: the depth model is one call in a pipeline, and the pipeline around it is where the cost and the fragility live. OrcaRouter puts 217 models behind a single OpenAI-compatible endpoint at 0% markup — the provider's list price passes through unchanged — and adds automatic failover across providers, so a depth stage that has to degrade gracefully can do so without a second vendor contract. Routing is not the interesting part of a depth pipeline, but it is the part that decides whether the pipeline stays up.
How to choose
Take UniDepth V2 when you want depth and geometry together, when you have a camera model you can supply, and when you need numbers you can cite. The ten-dataset evaluation is the strongest published evidence either model has, and the confidence output is genuinely useful for filtering bad predictions downstream.
Take Depth Pro 2 when you want a size dial and a video path. Starting with S on a laptop and moving to H+ on a server is a real workflow that Apple documents, and the separate focal-length estimator means you can run the camera solve once and reuse it. The temporal-consistency design only pays off on video, though — on stills you are paying for cross-frame attention you never use.
What you should not do is read a comparison of these two that ranks them. No shared benchmark exists, and the metrics each lab chose are the metrics each lab is good at. If you need a decision today, the tiebreakers are the output format and the license, not a winner.
