# Apple Depth Pro 2: Open Weights for Sharp, Video-Stable Metric Depth

> Source: <https://explainx.ai/blog/apple-depth-pro-2-open-weights-metric-video-depth-focal-length-research-license-2026?utm_source=rss&utm_medium=feed&utm_campaign=blog>
> Published: 2026-11-20 00:00:00+00:00

**Key takeaway:** Depth Pro 2 ships four open-weight metric depth models from 24M to 872M parameters plus a separate 23.5M-parameter focal-length estimator, so arbitrary footage without focal metadata can be converted to depth in meters, but the weights are non-commercial research only.

Apple has put **Depth Pro 2** on [Hugging Face](https://huggingface.co/apple/DepthPro2-focal?ref=explainx) and [GitHub](https://github.com/apple-aiml-research/ml-depth-pro-2?ref=explainx): four open-weight depth models, a small focal-length estimator, and a command line tool that turns photos, video and live camera feeds into metric depth maps. The catch is the license. The weights allow only non-commercial research.

The release is thin on published numbers. The [project page](https://apple-aiml-research.github.io/ml-depth-pro-2?ref=explainx) says Depth Pro 2 produces "sharp, metric, temporally consistent depth for images and video at full resolution." It does not give a runtime, a parameter-for-parameter comparison with the first Depth Pro, or a results table. This post covers what is confirmed, what is not, and how to try it.

## Quick reference: what Apple released

| Question | Answer (from the model card, README and project page) | 
|---|---|
| What does it do? | Metric monocular depth: depth in meters from one image, a clip or a live feed | 
| Which sizes? | S (24M parameters, 97 MB), B (97M, 386 MB), L (334M, 1.3 GB), H+ (872M, 3.5 GB) | 
| Default model | H+ | 
| Focal length | Separate estimator, 23.5M parameters (about 94 MB), for media with no known focal length | 
| Where does it run? | NVIDIA GPU (CUDA), Apple silicon GPU (MPS) or CPU | 
| Video? | Yes: buffered 16-frame clips with 4-frame overlap, or per-frame streaming | 
| License | Apple Machine Learning Research Model license: non-commercial research only | 
| Benchmarks published? | Qualitative claims only. No numbers on the pages we read | 
| Paper | Cited as a 2026 technical report; the arXiv ID is a placeholder | 

## What is monocular metric depth, and why does Apple's version matter?

A depth model looks at a single ordinary photo and guesses how far each pixel is from the camera. Two words in that sentence do most of the work.

**Monocular** means one camera and one frame, with no stereo pair and no lidar. The model has to infer distance from cues such as perspective, texture and the size of familiar objects.

**Metric** means the output is in real units. A relative depth map tells you the lamp is closer than the wall. A metric one tells you the lamp is 1.8 meters away. That difference decides whether the output is usable for measuring a room, placing virtual objects at the right scale or guiding a robot.

Metric depth from one image is harder than relative depth because the scale is ambiguous. A photo of a small model house and a real house can look alike. That is where the focal-length estimator comes in. Pixel measurements only turn into meters if you know the camera's focal length, and many photos and videos carry no reliable metadata. Apple ships a small separate network that predicts the focal length in pixels, so the depth model can work on arbitrary footage. The model card describes the estimator as a 23.5M-parameter component, loaded with `FOVEstimator.from_pretrained("apple/DepthPro2-focal")`.

## What is new in Depth Pro 2: video and sharper edges

The title of the report, as cited on the model card, is "Depth Pro 2: Sharp and Temporally Consistent Metric Monocular Depth." Those two adjectives are the pitch.

**Sharp** refers to edges. Depth maps from many models smear at object boundaries, which looks fine in a heat-map but ruins effects such as background blur or 3D parallax, where a halo around a person is obvious. The project page says Depth Pro 2 outperforms prior work on Boundary F1, a measure of how well predicted depth edges line up with real ones.

**Temporally consistent** refers to video. A single-frame model run on each frame of a clip gives depth that flickers, because small changes in the image produce different guesses. The project page says Depth Pro 2 beats prior work on Jitter, a measure of frame-to-frame instability. Both claims are averaged over the Sintel and Spring datasets, according to the page. It does not name the baselines or give the scores.

The architecture description is brief. A multi-view transformer processes one or more frames jointly. After the first third of its blocks, it alternates attention within each frame and across frames. A DPT decoder then predicts an inverse depth map for each frame. Put plainly, the model looks at several frames at once, so it can agree with itself over time instead of guessing frame by frame.

The video comparisons on the page use 2K to 4K sources. The H+ model runs at native resolution, and the baselines run at their default settings. That is a reasonable way to show sharpness, but it is also a presentation choice, not a controlled benchmark. Treat the claims as Apple's, until the paper and the numbers are public.

## The four sizes: which one should you run?

Laptop with a small chip, a fit for the 97 MB Apple Depth Pro 2 S model on Apple silicon

The README lists four depth models. The size spread is wide, so the choice matters.

| Model | Parameters | Download size | Good for | 
|---|---|---|---|
| S | 24M | 97 MB | Laptops, quick tests, live use (the README suggests S for laptops) | 
| B | 97M | 386 MB | A middle option when S looks too soft | 
| L | 334M | 1.3 GB | Better quality on a desktop GPU | 
| H+ | 872M | 3.5 GB | Default; best quality, native-resolution video | 

The README says the code runs in bfloat16 on NVIDIA and Apple GPUs and in float32 on the CPU. Core ML exports are on Hugging Face, and the repo includes a Mac live-depth demo that needs Apple silicon, macOS 15 or later and Xcode 16 or later.

If you only want to see whether the model works on your footage, start with S. If the edges look wrong, step up. Speed numbers are not published, so measure on your own hardware before you plan around live use.

## How to try Depth Pro 2 in five minutes

The README's quick start is short.

```
# install uv first: https://docs.astral.sh/uv/
git clone https://github.com/apple-aiml-research/ml-depth-pro-2
cd ml-depth-pro-2
uv run depth-pro-2 predict -i assets/example.jpg -o out/
```

The command writes a color depth PNG, or an MP4 for video. Options add 16-bit inverse-depth PNGs and NPZ files that hold depth in meters and the focal length in pixels. For code, the package has a Python API with `create_model_and_transforms`, `infer`, `infer_video` and `stream`.

On the Hugging Face side, the model card shows loading the focal-length estimator, moving it to a GPU, normalizing an image into a tensor and calling `focal_length_px`. For video, you pass frames shaped as (T, 3, H, W). The full depth model is `apple/DepthPro2-H`, and the whole family sits in the [apple/DepthPro2 collection](https://huggingface.co/apple/DepthPro2?ref=explainx).

One detail matters for video: a clip gets one focal length for all its frames. If your footage zooms, that assumption breaks, and the metric scale will drift.

## The license: research only

A box opening and releasing seeds toward open hands, picturing weights released for research use only

The weights ship under the Apple Machine Learning Research Model license. The README quotes it as permitting use "exclusively for Research Purposes," which it defines as non-commercial research and academic activities. The code has its own LICENSE file, but the weights are the valuable part, and they are the restricted part.

That has three practical effects.

- **No product use.** You cannot put these weights into a shipped app, a paid API or an internal tool that supports revenue.
- **Fine for study.** Students, labs and hobby research on non-commercial terms are the intended audience.
- **Training and evaluation are absent.** The repo has inference code and weight links, not training or evaluation code, so you cannot reproduce the results or retrain on your own data.

Apple has used the same research-only approach before. For another recent open release with similar terms, see our coverage of [Apple FastFrames-L, the open-weights text-to-video model](https://explainx.ai/blog/apple-fastframes-l-interactive-text-to-video-open-weights-2026). The pattern is consistent: Apple publishes weights so researchers can inspect and build on them, and keeps commercial use for its own products.

## What is still missing

Several things you would want before relying on this are not in the sources.

1. **No release date.** The model card and README show a 2026 copyright, but no day.
2. **No numbers.** The project page claims wins on Boundary F1 and Jitter and does not print the scores.
3. **No speed figures.** There is no frames-per-second table for any model size or device.
4. **No Depth Pro 1 comparison.** The page does not say how much better Depth Pro 2 is than its predecessor.
5. **A placeholder paper link.** The arXiv ID reads as a template, so the technical report is not yet citable.
6. **A very new repo.** At the time we checked, the GitHub page showed no stars, watchers or forks, so there is no community testing yet.

None of this means the model is weak. It means the evidence so far is Apple's own gallery. When the paper lands, check the tables for the baselines and the exact datasets.

## How does it fit with other depth and 3D work?

Depth is a building block, and a few recent stories show where it goes.

- **Camera tracking on long video.** NVIDIA's work on[metric depth and camera motion for long video](https://explainx.ai/blog/nvidia-metrictrail-camera-motion-metric-depth-long-video-dpvo-moge-2-dinov3-kitti-ate-2026) pairs depth with pose estimation. Video-stable depth is exactly the input that kind of system wants.
- **Local models on Mac.** If you run models on Apple hardware, our[MLX guide for Apple silicon](https://explainx.ai/blog/mlx-apple-silicon-run-llms-complete-guide-2026) covers the runtime side, and the Depth Pro 2 Core ML exports suit the same machines.
- **Apple's agent research.** Apple also publishes work on[an online-RL agent that plans and grounds actions in live iOS apps](https://explainx.ai/blog/apple-online-rl-planning-grounding-agent-live-ios-apps-8b-gigppo-2026) and[faster retrieval for RAG](https://explainx.ai/blog/apple-amortized-mips-supportnet-keynet-faster-rag-retrieval-icml-2026-natural-questions-1-9x-faster-2026) . Depth Pro 2 shows that the research arm covers vision as well as language.

For builders, the realistic uses today are prototypes: portrait-style blur on video, rough scene scale for AR sketches, or a depth input for a robotics or mapping experiment. Any of these must stay non-commercial until Apple says otherwise.

## What should builders do now?

- **Try S first.** It is 97 MB and the README suggests it for laptops.
- **Test your own footage.** Judge edges and flicker on your clips, not Apple's gallery.
- **Keep the focal-length assumption in mind.** One focal length per video clip fits fixed-lens footage, not zooms.
- **Do not ship it.** Read the license before copying any weight into a product.
- **Watch for the paper.** When the arXiv link goes live, compare the Boundary F1 and Jitter tables with other open depth models.

## Frequently asked questions

### Is Depth Pro 2 open source?

Not in the usual sense. The weights are public, and the code is on GitHub, but the weights license limits use to non-commercial research. "Open weights, research license" is the accurate label.

### Does Depth Pro 2 need a GPU?

No. The README says it runs on an NVIDIA GPU, an Apple silicon GPU or the CPU. It uses bfloat16 on GPUs and float32 on CPU. Expect CPU runs to be slow on the larger models, and use S on a laptop.

### What does the focal-length estimator do?

It predicts the camera focal length in pixels from the image or video frames, so the depth model can output meters when the file has no camera metadata. It is a 23.5M-parameter model on Hugging Face.

## Related reading

Official sources: the [Depth Pro 2 project page](https://apple-aiml-research.github.io/ml-depth-pro-2?ref=explainx), the [GitHub repository](https://github.com/apple-aiml-research/ml-depth-pro-2?ref=explainx) and the [Hugging Face collection](https://huggingface.co/apple/DepthPro2?ref=explainx).

*Details on this page are accurate as of November 20, 2026. Model sizes, license terms and the paper link may change as Apple updates the release.*
