
Depth Pro 2 vs Depth Pro: Sharpness Held, Jitter Fixed
- 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 original Depth Pro has an unusual problem as a baseline: the checkpoint Apple published is not the model the paper describes. The model card says so directly — the checkpoint in the repository is a reference implementation which has been re-trained, and its performance is close to the model reported in the paper but does not match it exactly. So any comparison between Depth Pro 2 and Depth Pro is a comparison against a stand-in, including the one on Depth Pro 2's own card.
With that caveat stated, the shape of the upgrade is clear enough. Depth Pro 2 keeps the property that made Depth Pro notable — sharp depth at object boundaries, in metres, from a single image — and adds a second one: depth that holds steady from frame to frame. The paper title says it plainly. Depth Pro was Sharp Monocular Metric Depth in Less Than a Second. Depth Pro 2 is Sharp and Temporally Consistent Metric Monocular Depth. Everything that changed between them serves the second adjective; almost nothing changed about the first.
Two years, two very different releases
Depth Pro landed on 3 October 2024 with a paper, arXiv 2410.02073, and a stated result: a 2.25-megapixel depth map in 0.3 seconds on a standard GPU. The repository carries 568 likes and just over 6,200 downloads. It shipped as one checkpoint, depth_pro.pt, and its card describes the technical contributions — an efficient multi-scale vision transformer for dense prediction, a training protocol that mixes real and synthetic data, dedicated boundary-accuracy metrics, and state-of-the-art focal length estimation from a single image.
Depth Pro 2's repositories were created in November 2026: the card on 3 November, the weights on 5 November, Core ML exports on 9 November, the focal estimator and the code repository on 19 November. The card and the GitHub README both point at an arXiv identifier of 261X.XXXXX, which is a placeholder rather than a preprint. There is no research page for it on machinelearning.apple.com — that path returns 404 — and no benchmark numbers anywhere outside the card's own scatter plot.
So the honest framing is that Depth Pro 2 is a weights release with documentation, not a publication. Depth Pro was a paper with a checkpoint. If you are citing something, one of these gives you a number and the other gives you a repository.
What actually changed in the architecture
Depth Pro was a single-image model. Depth Pro 2 is a multi-view transformer that processes one or more frames jointly, and the mechanism is specific: after the first third of its blocks, attention alternates between within each frame and across frames. A DPT decoder then predicts an inverse depth map per frame, assembling tokens from several blocks into feature maps at one thirty-second to one quarter of the input resolution.

That is the entire architectural argument for the sequel, and it has a visible consequence in how you run the model. Depth Pro 2 has two video modes. Buffered, the default for files, processes clips of 16 frames overlapping by four, which keeps depth steady where clips meet. Unbuffered predicts each frame as it arrives, from that frame and the 15 before it, and is the mode for a live camera. Apple documents the cost of the live path rather than hiding it: the first frames of a stream have fewer predecessors to attend to, so their depth is less steady.
Depth Pro has neither mode, because it has no cross-frame attention to run. Feed it a video and you get independent per-frame estimates, and any flicker between them is yours to smooth. For stills that is not a loss; for anything with motion in it, it is the difference the sequel was built to close.
What did not change
The metric conventions are identical, which is good news for anyone migrating: the two models agree on what the numbers mean. Depth is the distance along the camera's viewing axis — the z coordinate — not the straight-line distance from the lens. Inverse depth in Python is normalized by focal length, equal to focallength_px / (width * depth). Depth scales with focal length, so a focal length 10 percent too long places everything 10 percent too far away. Sky pixels come back as infinite depth or, especially on the smaller models, as a finite depth of several kilometres, and the documentation tells you to clip the range before making a point cloud.
Focal length estimation is still a first-class part of the package, and still skippable. Photos that record a 35mm-equivalent focal length in EXIF use it; everything else goes through the estimator. Depth Pro 2 splits that estimator into a separate 23.5M-parameter repository, apple/DepthPro2-focal, which loads alongside any size and can also be run on its own — useful if you want the camera solve once and the depth model many times. With Depth Pro, the same capability was baked into the single checkpoint with no way to separate the two.
Core ML support carries over and improves. Depth Pro 2 ships exported Core ML packages in a dedicated repository and includes a minimal Mac demo app that shows live depth from the camera, requiring Apple silicon, macOS 15 and Xcode 16. That is a lower bar to a working demo than the original offered.
Four sizes against one checkpoint
The biggest practical difference is that Depth Pro 2 refuses to be a single model. Apple publishes four, with parameters and download sizes:
• S — 24M parameters, 97 MB, the recommended starting point on a laptop.
• B — 97M parameters, 386 MB.
• L — 334M parameters, 1.3 GB.
• H+ — 872M parameters, 3.5 GB, the default the command line downloads.
The size you pick changes the licence not at all and the behaviour quite a lot, since the smaller models are documented as producing more of those spurious multi-kilometre sky depths. H+ is the default for a reason, but on a laptop S is the difference between working and not.
Depth Pro is one checkpoint with no dial. That is not a criticism — it was a 2024 release and the paper reported a single 0.3-second result — but it does mean the migration question "which size" has no counterpart on the old side, and any comparison that treats "Depth Pro" as one row against four is flattening a real decision.
The numbers, such as they are
Depth Pro 2's card reports Boundary F1 and Jitter averaged over the Sintel and Spring datasets, claims to outperform prior work operating on a single frame or multiple frames on both, and plots the comparison. Depth Pro is one of the markers on that plot. The axes are labelled — Boundary F1 from 0 to 0.3, jitter from 0.1 down to 0 — but there is no per-model table, so the improvement is readable to about a hundredth: H+ sits near 0.29 Boundary F1 at roughly 0.017 jitter, the original Depth Pro at roughly 0.235 and 0.095.

Treat that as a vendor-reported, unreproduced claim, which it is. No third party has published a Boundary F1 or Jitter number for Depth Pro 2, and the technical report that would define the evaluation protocol is a placeholder. The original Depth Pro, by contrast, ships the boundary metrics in its own codebase — eval/boundary_metrics.py exposes SI_boundary_F1 and SI_boundary_Recall — so if you want to run the comparison yourself, the tooling for half of it is public and the target checkpoint is a re-trained approximation of a paper model.
Licence and the commercial question
Unchanged, and worth stating because it is the kind of thing that gets missed in an upgrade. Both models' weights are under the Apple Machine Learning Research Model licence, which the repository defines as permitting use exclusively for research purposes, meaning non-commercial scientific research and academic development. The code is separately licensed, and Depth Pro 2's acknowledgements credit TIMM and DINOv3 as its foundations.
So the upgrade is free in the licence sense and unavailable in the product sense at both ends. If you are shipping for revenue, neither checkpoint is the one, and the conversation you need is with Apple rather than with a licence file.
Migrating, concretely
Depth Pro 2 is not a drop-in replacement for Depth Pro, because the package name, the entry point and the model identifiers all changed. Depth Pro installs from apple/ml-depth-pro and runs as depth-pro-run -i ./data/example.jpg. Depth Pro 2 installs from apple-aiml-research/ml-depth-pro-2 and runs as uv run depth-pro-2 predict -i assets/example.jpg -o out/, with -m depthpro2-s through -m depthpro2-h for size. In Python the import is depth_pro_2 and the construction path is create_model_and_transforms or DepthPro2.from_pretrained.
The output surface changed shape too. Depth Pro's model.infer() returns a dictionary with depth and focallength_px. Depth Pro 2's infer() returns depth, and the focal length is more naturally obtained from the separate estimator or supplied as f_px when you have it. If you have a pipeline that reaches into the prediction dict, budget an afternoon.
What you get for the migration is a size dial, a video path with two documented modes, a separated focal estimator, and Core ML packaging with a working demo app. What you do not get is a paper, a per-model benchmark table, or any independent confirmation that the new model is better — only a scatter with labelled axes and no table behind it.
None of this is routed through us, and it should not be described as though it were: Depth Pro 2 is research-licensed weights from Apple's own repositories. The place a routing layer earns its keep is one step out — the scene description, the decision, the write-up around your depth stage are language-model calls, and OrcaRouter puts 200-plus hosted models on one key at 0% markup over provider list price, with automatic failover across providers, so a research checkpoint with no release cadence is the only fragile thing in the pipeline rather than one of several.
If you are on Depth Pro and it works, the reason to move is video. If you are on Depth Pro and you only ever run stills, the sharpness you have is the sharpness the sequel claims to have kept, and the honest thing to do is wait for the technical report before you spend an afternoon on an import rename.

