Apple introduced Depth Pro 2 in a preprint submitted on November 19. The authors describe one model for metric depth from images and video that combines sharp boundaries with temporal consistency. These are reported research results, rather than independent confirmation by SIExpo. The released weights are restricted to noncommercial research.
Why video needs a separate assessment
A convincing individual frame does not establish that distance estimates remain consistent across frames. For an application using video, usefulness also depends on that behavior over time. Our interpretation is that model selection should follow the required consistency of its outputs, rather than the most attractive demonstration image.
One proposed acceptance example is a stationary object recorded with gentle camera motion. A team could inspect both the object's boundary and its estimated distance. We have not performed that experiment: the scenario illustrates how two requirements can be examined separately. An application relying on absolute distances would also need appropriate reference measurements, rather than visual inspection alone.
The released implementation has conditions
The official repository provides inference code and models, but not training or evaluation code. Its documentation calls for a new stream when changing the video, camera or frame size. The Mac demonstration requires Apple silicon, macOS 15 or later, and Xcode 16 or the corresponding command-line tools.
Integration therefore needs explicit management of stream state. A useful design record would associate each session with its camera and input format. Otherwise, incorrect interface use can be confused with a weakness in the model. This recommendation does not replace the documentation for the specific implementation version being evaluated.
Code does not automatically license the weights
The model license defines research use and excludes commercial exploitation and product development. Code has a separate license, while third-party dependencies are identified independently. Downloading the repository consequently does not provide sufficient grounds for embedding the weights in a commercial service. The intended use needs to be checked before making a product commitment.
For a research pilot, we would retain the exact model version, permitted purpose and assessment procedure. A dataset with motion, varied distances and thin boundaries could provide more useful evidence than one appealing example. Processing time, resource use and output quality should be reported separately, with failures preserved alongside successful cases. Depth Pro 2 offers a new subject for that investigation; suitability for a particular system remains a question of testing and permitted use.



