
Depth Pro 2 vs Depth Anything 3: Metres or Camera Poses
- 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
Ask both models the same question — how far away is this? — and only one of them answers in metres. Depth Pro 2, Apple's metric monocular depth model, returns a depth map in metres along the camera's viewing axis, plus the focal length it used. Depth Anything 3, the ByteDance Seed foundation model released in November 2025, lists "relative depth" as its first capability, alongside camera pose estimation and pose conditioning. It gives you depth up to an unknown scale, and it gives you the cameras that make that scale recoverable.
That single difference decides most integrations before accuracy enters the conversation, and it is the reason a straight benchmark comparison between these two is less useful than it looks. The two labs report on different metrics, against different baselines, and neither has run the other's evaluation. What follows is what each repository actually documents, with the vendor claims labelled as such.
Two different outputs, one letter apart
Depth Anything 3's prediction object is the clearest place to start, because it is unusually explicit. A single call returns depth as [N, H, W] float32, a confidence map of the same shape, extrinsics as [N, 3, 4] world-to-camera matrices, and intrinsics as [N, 3, 3]. Exports go to GLB, NPZ, PLY, mini-NPZ, Gaussian-splat PLY or a Gaussian-splat video.
Depth Pro 2's output is narrower and more immediately useful if the answer you need is a number. It returns a depth map in metres, a focal length in pixels, and — in the command line's NPZ output — those two arrays and nothing else. Conversion to a point cloud is left to you, using the focal length and the image centre as principal point, or your real principal point if the camera is calibrated.
• Scale — Depth Pro 2's output is metric; Depth Anything 3's depth is relative, with scale recovered from the pose and intrinsics it also returns.
• Camera parameters — Depth Anything 3 predicts extrinsics and intrinsics; Depth Pro 2 predicts focal length only, and has no notion of where the camera is.
• Confidence — Depth Anything 3 returns a per-pixel confidence map; Depth Pro 2 returns none.
• Input — Depth Pro 2 takes one photo, a video file, or a live camera feed; Depth Anything 3 takes a list of images as one view set.
• Licence — Depth Pro 2 weights are research-only under the Apple Machine Learning Research Model licence; Depth Anything 3 is Apache 2.0 at 0.08B and 0.12B, CC BY-NC 4.0 at 0.35B and 1.15B.
The four sizes do not line up
Both families publish four checkpoints, and the parameter counts land in almost the same places, which makes it tempting to compare them one to one. Resist that.
Depth Anything 3 publishes 0.08B, 0.12B, 0.35B and 1.15B — 34.3M, 135.4M, 410.9M and 1,355.7M parameters according to the safetensors metadata. Depth Pro 2 publishes 24M, 97M, 334M and 872M. The two smallest Depth Pro 2 sizes are meaningfully smaller than their Depth Anything 3 counterparts, and the largest is meaningfully smaller than DA3 Giant.
The licences split along the same line as the sizes, and this is the detail that matters commercially: Depth Anything 3's two small checkpoints are Apache 2.0, and its two large ones are CC BY-NC 4.0. Depth Pro 2's four are all under the same research-only licence, defined in the repository as non-commercial scientific research and academic development. So if you need a commercially licensable depth model at all, Depth Anything 3-Small or -Base is the only option in this comparison, and Depth Pro 2 is not one at any size.
Depth Anything 3 also has a 1.1 revision, DA3-GIANT-1.1 and DA3-LARGE-1.1, with identical parameter counts to the originals and repositories created on 11 December 2025. The original GIANT card describes 3D Gaussian estimation as an additional capability the large sizes carry, so the export surface is wider than depth alone.
What each lab claims, and against what
Depth Anything 3's card makes two comparative claims and names the baselines for both: it significantly outperforms Depth Anything 2 for monocular depth estimation, and VGGT for multi-view depth and pose estimation. Those are specific, falsifiable statements about named models, and the detailed numbers live in the paper rather than the card.
Depth Pro 2's card makes one claim and names its baseline only as "prior work operating on a single frame or multiple frames". It reports that Depth Pro 2 outperforms that set on Boundary F1 and Jitter, averaged over the Sintel and Spring datasets, and illustrates it with a scatter plot of those two metrics, its axes running Boundary F1 from 0 to 0.3 and Jitter from 0.1 down to 0.

Depth Anything 3 is in that chart — as DA3-NGL, one of the open-circle multi-frame baselines sitting at roughly 0.06 Boundary F1 and 0.018 jitter. So a comparison exists, but only on Apple's two metrics, averaged over Sintel and Spring, with no Depth Anything 3 depth numbers printed alongside. ByteDance's own comparison set is Depth Anything 2 and VGGT, on the metrics ByteDance chose, and DA3-NGL is not discussed there. The two sets do not meet: whichever card you read, the other lab chose the axes, and neither published a shared table.
That is not a dodge around a winner. It is the actual state of the evidence, and it is worth stating plainly because the alternative — picking whichever vendor chart flatters the model you already prefer — is how bad procurement decisions get made.
Video consistency is Apple's whole argument
Depth Pro 2 exists because of a specific weakness in monocular depth: run a per-frame model on video and the depth flickers, because each frame is estimated without knowledge of its neighbours. Apple's fix is architectural. The model is a multi-view transformer over the input frames, and after the first third of its blocks, attention alternates between within-frame and across-frame. A per-frame DPT decoder then assembles tokens from several blocks into feature maps at one thirty-second to one quarter of the input resolution.

The operating modes follow from that. Buffered mode, the default for video files, processes clips of 16 frames overlapping by four so depth stays steady across clip boundaries. Unbuffered mode predicts each frame as it arrives from that frame and the 15 before it, for camera use, with the documented caveat that the first frames of a stream have fewer predecessors and therefore less steady depth. A video gets one focal length across all frames — the median of the estimates over the first window.
Depth Anything 3 handles multiple views jointly by design, since that is what a view-set model does, but its framing is reconstruction rather than stabilisation. The card's language is about recovering visual space from any views, and the outputs — extrinsics, intrinsics, point maps, Gaussian splats — are the language of a 3D reconstruction pipeline. If your problem is "this video flickers", Depth Pro 2 has a mode for it and Depth Anything 3 has a different product.
Getting depth into metres either way
Depth Anything 3's relative depth is not a limitation so much as a division of labour: the model hands you the cameras, and the cameras are what turn relative depth into metres. That conversion is real work — a scale factor per scene, from a known baseline, a known object size, or a sensor — and it is the same work Depth Pro 2 has already done internally.
Depth Pro 2's metric output carries its own failure modes, and they are documented. Depth scales with focal length, so a focal length 10 percent too long puts 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, so the README tells you to clip the range before building a point cloud or a colour image. In Python, inverse depth is normalized by focal length and equals focallength_px / (width * depth). A stream belongs to one video and keeps its recent frames and its focal length until you open a new one.
Neither model replaces a depth sensor on a textureless or reflective surface. Both are inferring scale from visual evidence, and both will fail in the same kinds of places for the same kinds of reasons.
What it costs to run both
The practical difference is in the dependency graph. Depth Pro 2 installs as a Python package from the GitHub repository with uv, exposes a command line (uv run depth-pro-2 predict -i photo.jpg -o out/) and a Python API, and ships Core ML exports for Apple silicon plus a Mac demo app. The repository is explicit that it contains code to run the models but not training or evaluation code.
Depth Anything 3 installs as an editable package from the ByteDance Seed repository and exposes a CLI (da3 auto path/to/images --export-format glb) with an optional backend process for repeated inference, which is the kind of detail that only appears when someone is running it in a loop.
Both are single-purpose research checkpoints, and neither is something you route. What sits around a depth stage is language-model work — describing a scene, deciding what to do about it, writing the result down — and that is where a single endpoint earns its keep: OrcaRouter puts 200-plus hosted models behind one OpenAI-compatible API at 0% markup over provider list price, so the reasoning stages around your depth model are one key and one bill. The depth checkpoint stays a dependency you pin; the pipeline around it does not have to be.
Which one to start with
Start with Depth Anything 3 if you are reconstructing a scene from several views and you want the cameras, the point maps and the Gaussian splats in one call, and if an Apache 2.0 licence at the small sizes is what makes the project possible at all. Its comparison set is named, its outputs are wide, and the small checkpoints are the only commercially licensable things in this article.
Start with Depth Pro 2 if the answer you need is a distance in metres from a single frame or a live feed, and if frame-to-frame steadiness is the property your application fails without. Go in expecting to clip sky depth, to pin the focal length deliberately, and to be early — the repositories have downloads in the tens, the code has one commit, and the technical report is still a placeholder.
What neither model gives you is a defensible claim that it beats the other, because neither lab ran the test. Pick on output format, licence and whether your input is a video. Those three questions have answers today; the leaderboard does not.

