> ## Content Index
> Fetch the complete content index at: https://www.notatechguy.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Apple Depth Pro 2 claims sharp, steady video depth
- URL: https://www.notatechguy.com/apple-depth-pro-2-claims-sharp-steady-video-depth/
- Published: 2026-11-22T18:35:50.000Z
- Updated: 2026-11-22T18:35:50.000Z
- Description: Apple researchers released Depth Pro 2 with open code and weights, claiming metric depth that is both sharp and temporally consistent for video.
- Author: Marcello Babbili
- Tags: Technology & AI, OpenAI

Apple released Depth Pro 2 on November 19, claiming one network produces video depth that is both sharp and steady, breaking a tradeoff that has split the field for years [S¹](https://arxiv.org/abs/2611.26603v1?ref=notatechguy.com).

Until now, multi-frame methods kept depth stable across video but blurred object edges, while single-frame methods sharpened boundaries but flickered frame to frame [S¹](https://arxiv.org/abs/2611.26603v1?ref=notatechguy.com). The team, led by Peiyun Hu and Vladlen Koltun at Apple, put code and weights on GitHub and Hugging Face [P⁵](http://github.com/apple-aiml-research/ml-depth-pro-2?ref=notatechguy.com)[P³](https://huggingface.co/apple/DepthPro2?ref=notatechguy.com).

**My read:** The claim is bold but the evidence is thin in one specific way. The paper is a preprint and has not been peer-reviewed [S¹](https://arxiv.org/abs/2611.26603v1?ref=notatechguy.com). It states that Depth Pro 2 outperforms prior work across all measured dimensions [S¹](https://arxiv.org/abs/2611.26603v1?ref=notatechguy.com). These figures are Apple's own; no error bars, released-code reproductions, or third-party benchmarks have been published to check them [S¹](https://arxiv.org/abs/2611.26603v1?ref=notatechguy.com). Until independent teams reproduce the results, this is Apple grading its own homework. The number I would watch is any third-party benchmark on standard depth datasets like KITTI or NYU Depth, which the community uses to settle exactly these claims.

### Multi-frame depth stays steady but blurs edges

Monocular depth estimation, the task of guessing how far away each pixel is from a single camera, has long forced researchers to pick a side. Multi-frame methods, which look at several video frames at once, produce depth that stays stable from frame to frame but blurs the edges where objects meet. Single-frame methods sharpen those boundaries but flicker when applied to video, because each frame is judged in isolation [S¹](https://arxiv.org/abs/2611.26603v1?ref=notatechguy.com).

The result is a familiar problem for anyone building augmented reality, autonomous driving systems, or 3D video effects. A depth map that flickers makes virtual objects placed in a scene appear to jitter. A depth map that blurs edges makes thin structures like chair legs or power lines dissolve into the background.

Depth Pro 2's pitch is that one network can do both jobs at once.

It claims to produce depth maps that are metrically accurate, meaning distances are in real-world units like metres, sharp at object boundaries, and steady across frames [S¹](https://arxiv.org/abs/2611.26603v1?ref=notatechguy.com)[P³](https://huggingface.co/apple/DepthPro2?ref=notatechguy.com).

### A single network predicts depth at native resolution for images and video

The architecture is a single scalable network that processes both images and video and predicts depth at the input's native resolution [S¹](https://arxiv.org/abs/2611.26603v1?ref=notatechguy.com). The Hugging Face model card describes it working from a single photo, a whole video, or a live feed [P³](https://huggingface.co/apple/DepthPro2?ref=notatechguy.com). The team released a family of models in different sizes, with larger versions producing sharper and steadier depth, according to the model documentation [P³](https://huggingface.co/apple/DepthPro2?ref=notatechguy.com).

The GitHub repository, created on November 3 under the apple-aiml-research organisation, had 437 stars by the time of writing [P⁵](http://github.com/apple-aiml-research/ml-depth-pro-2?ref=notatechguy.com). The code is in Python and the licence is listed as "Other" [P⁵](http://github.com/apple-aiml-research/ml-depth-pro-2?ref=notatechguy.com).

Open-weight releases from major labs set the baseline that smaller teams build on. When Apple puts both code and weights out, independent researchers can run the model on their own data and publish their own numbers, which is the only way to close the gap between a vendor's self-reported results and verified performance. That same gap, between self-reported and independently checked, is what [OpenAI tried to address with its perplexity-based safety metric proposal](https://www.notatechguy.com/openai-publishes-perplexity-scores-proposes-shared-ai-safety-metric/) earlier this month.

### No independent benchmark has checked the Pareto claim

The paper states that Depth Pro 2 defines a new Pareto frontier in metric accuracy, sharpness, temporal consistency, and runtime [S¹](https://arxiv.org/abs/2611.26603v1?ref=notatechguy.com). A Pareto frontier, in this context, means no other model beats it on any one of those measures without sacrificing another.

That is a strong claim. No independent third party has benchmarked Depth Pro 2, and the paper's figures are Apple's own, drawn from internal evaluations with no published error margins [S¹](https://arxiv.org/abs/2611.26603v1?ref=notatechguy.com). The gap between a lab's own benchmarks and independent verification is the same problem [OpenAI's perplexity score proposal](https://www.notatechguy.com/openai-publishes-perplexity-scores-proposes-shared-ai-safety-metric/) tried to close for safety metrics, and it remains open here.

For a team building an augmented reality app, the practical first step is to pull the weights from Hugging Face, run them on a short video clip with known depth (a scene filmed with a depth sensor, for instance), and compare the output frame to frame. An AR developer who currently runs a single-frame model for edge sharpness plus a separate multi-frame pass for temporal stability can test whether Depth Pro 2 replaces both steps with one inference call [S¹](https://arxiv.org/abs/2611.26603v1?ref=notatechguy.com). If the edges stay sharp and the depth values do not flicker, the claim holds for that use case. If they do not, the tradeoff the paper claims to have broken may still apply.

The next checkpoint is whether independent benchmarks appear on standard depth estimation datasets. The open code makes that test possible. The preprint status means it has not happened yet.

---

*Sources: [S1 — Depth Pro 2: Sharp and Temporally Consistent Metric Monocular Depth](https://arxiv.org/abs/2611.26603v1?ref=notatechguy.com) · [P2 — Depth Pro 2: Sharp and Temporally Consistent Metric Monocular Depth - ](https://machinelearning.apple.com/research/depth-pro-2-monocular?ref=notatechguy.com) · [P3 — apple/DepthPro2 · Hugging Face](https://huggingface.co/apple/DepthPro2?ref=notatechguy.com) · [P4 — Depth Pro 2: Sharp and Temporally Consistent Metric Monocular Depth](https://arxiv.org/abs/2611.26603?ref=notatechguy.com) · [P5 — apple-aiml-research/ml-depth-pro-2](http://github.com/apple-aiml-research/ml-depth-pro-2?ref=notatechguy.com)*

---

*Written from 5 sourced items, 4 of them primary.*