A meerkat’s fur exposes the problem facing video creators in Apple’s Depth Pro 2 paper, submitted on 19 November 2026: a sharp outline is only half the job. For depth-driven blur, fine edges and consistent depth need to survive together as the scene moves.
In the authors’ comparison of distance estimates across frames, one rival preserves the fur but puts it at inconsistent depths; another stays steadier while softening the outline. Apple presents Depth Pro 2 as a way to retain both detail and continuity. Fur is an unforgiving place to hide that trade-off.
This is a research release, not an unrestricted editing tool. Apple’s model licence permits non-commercial scientific research and academic development, while explicitly excluding product development and commercial exploitation. Downloadable weights do not make it a ready-to-use option for paid client work.
Why a sharp still frame is not enough
A depth map assigns a distance estimate to each pixel. Here, that means distance along the camera’s viewing direction, rather than a straight line from the camera to each point. Apple’s research release estimates those distances from ordinary photos or video, without requiring a separate depth camera.
For a depth-driven effect, those estimates help decide how much to blur different parts of the picture. Misplace a fine boundary and foreground detail can blend into the background. Let the estimated distance jump between frames and the effect can change even when the scene does not justify it.
The practical implication is that consistency should not mean freezing the map. If a subject walks towards the camera, its estimated depth should change. The useful question is whether that change follows the movement, while keeping the subject’s outline intact.
Sharp single-image depth is not new: the original Depth Pro paper was first submitted in October 2024. The new emphasis is preserving detail across moving footage. Unlike movable layers in AI-generated designs, a depth pass does not provide separate objects to rearrange; it supplies distance information for an effect.
What Apple’s blur results actually establish
Apple’s application evaluation feeds different depth estimates into the same fixed BokehMe blur renderer across 220 RealBokeh still scenes. Its largest H+ model leads the five reported measures. A separate 197-frame handheld clip provides a qualitative video example.
Holding the renderer fixed helps isolate the contribution of the depth estimate. But the still-image scores cannot establish temporal stability, and one video example cannot settle how reliably the effect works across an editor’s footage. Independent replication of these findings remains unestablished.
Nor does Apple win every comparison: in the paper’s depth benchmarks, InfiniDepth scores higher on Sintel video-boundary accuracy, while VeloDepth records less jitter on Spring. Those exceptions matter more than a sweeping claim that flicker has been solved.
For creators, the distinction is between improving an input to an effect and validating a finished editing workflow. That also separates this work from ReTiNA’s recorded-brainwave film effects: ReTiNA explores how signals steer visual style; Depth Pro 2 estimates the geometry of existing footage.
Should footage be processed live or in clips?
The repository’s recorded-video defaults process 16-frame clips with four-frame overlaps. Live mode instead uses the arriving frame and up to 15 previous frames, allowing an estimate without waiting for future footage.
Live mode is not automatically a lower-quality choice: Apple reports broadly similar quality between its default modes on 300-frame, 720p DynamicReplica sequences with known camera calibration. That is a controlled result, not a guarantee for every camera feed.
Choose the mode around when the result is needed, then consider model size, resolution and how many frames it uses together. Processing recorded footage can tolerate a wait; a live effect has a deadline for every frame. Neither choice removes the need to inspect the final moving image.
What a Mac experiment actually needs
Apple lists four downloadable model sizes, from a 97 MB small model to a 3.5 GB H+ model, plus a shared 94 MB focal-length estimator. These are download sizes, not the memory an experiment will consume.
The minimal Mac demo requires Apple silicon, macOS 15 or later, and Xcode 16 or its command-line tools. Apple documents roughly 1.5 GB of memory for the small model at 854×480, rising to about 5 GB at 1280×720.
Its four-frame demo timings on an M4 Max with 64 GB report 39 frames per second at 854×480 and 11 at 1280×720 for that small model. Those are model-prediction timings: preprocessing, rendering, initial compilation and warm-up are excluded. They are not full-app performance guarantees. The demo’s timing notes give the conditions.
The paper’s 53-fps result at 1080p uses the small model on an NVIDIA B200, with focal length supplied. It does not establish phone or ordinary-laptop performance. For a permitted experiment, budget for the actual resolution and complete processing chain, not just a small download.
Check the depth pass, not just the colour preview
The documented exports include coloured previews, 16-bit inverse-depth images for compositing, and numeric depth arrays. Two details are worth checking: a focal-length estimate 10% too high scales depths 10% too far; the 16-bit video export uses the first frame’s range, so closer points later can saturate at white.
That suggests two different diagnostic questions. Is the model estimating the scene incorrectly, or is the chosen export losing information? Compare the numeric output before blaming an apparent failure on the model. A detailed, steady colour preview is not a tape measure.
For research within the licence, a useful next experiment would examine fine moving boundaries, genuine changes in distance and the rendered blur together. The released repository includes inference code, but not the training or evaluation code; it does not supply the complete published benchmark suite.
For a commercial editor or product prototype, the model-use restriction comes first: even an unpaid product-development project falls outside the stated scope. The broader creative lesson remains useful regardless of adoption. Judge video depth in motion, where both the outline and the distance estimate have to keep making sense.
