Depth Anything 3 vs Depth Pro: the metric depth matchup

Depth Anything 3 and Apple Depth Pro are the two names that come up when one image must yield real distances. The one shared benchmark table, printed in the DA3 paper, puts the DA3 metric tier ahead of Depth Pro on all five test sets, read 28 October 2026. Depth Pro keeps two edges that the table never scores. Its boundaries come out sharper, at higher native detail. And it guesses the focal length from the image alone. For business use, the licenses decide first. The DA3 metric tier ships under Apache 2.0. Depth Pro weights are locked to research use.

Last verified: 2026-10-28

Two metric depth models, two years apart

Depth Pro came first. Apple released it in October 2024. The paper, Depth Pro: Sharp Monocular Metric Depth in Less Than a Second, was published at ICLR 2025. It is one model with one job. Read a single image. Return depth in meters. Do it with no camera data at all. No focal length, no lens hints, no distance clues. The paper claims a 2.25-megapixel depth map in 0.3 seconds on a standard GPU.

Depth Anything 3, or DA3, is the younger system. The ByteDance Seed team released it in November 2025, paper here. DA3 is not one model. The official repository ships a family. A main series in four sizes predicts relative depth. A monocular series does the same job at higher quality. One metric tier, called DA3METRIC-LARGE, predicts meters. Relative depth orders a scene from near to far. Metric depth measures meters. That difference is the whole reason this comparison exists.

For metric work, the honest matchup is Depth Pro against the DA3 metric tier, not the family name. The DA3 main series does not compete here. Its output is relative. Its story against Depth Anything V2 lives on the DA3 vs DA2 page.

The two designs share an ancestor. Depth Pro splits the image into patches at several scales. A shared DINOv2 encoder reads every patch. A DPT-style decoder then fuses the patch features into one map. DA3 runs a plain DINOv2 transformer as its backbone. Same encoder family, two very different builds on top of it.

The only table that runs both models

No neutral lab has run these two side by side. One table exists, and it sits in the DA3 paper itself, section 7.4. The DA3 team scored five metric test sets under one protocol. Both models got the same treatment. These are their numbers, read 28 October 2026. The metric is delta1 accuracy, the share of pixels within 25 percent of the true depth. Higher is better.

DA3-metric wins all five. The ETH3D row is the widest gap, 0.917 against 0.386. On error rate the same set reads 0.104 against 0.349. The DA3 paper calls ETH3D its state-of-the-art set and names UniDepthv2, not Depth Pro, as the second-best system there, at 0.863.

Read the table with three cautions.

First, one team scored both models. The DA3 authors ran Depth Pro under their own protocol. Depth Pro has published no reply table. Its paper also predates DA3 by thirteen months.

Second, the Depth Pro paper reports different numbers for itself, on its own six-set suite. It prints 41.5 on ETH3D there, and 89.0 on Sun-RGBD. The protocols differ. The two papers never form one clean column. The DA3 table is the only place both models sit under one scoring rule.

Third, the shipped Depth Pro weights are not the paper model. The code repository and the Hub card both say so. The released checkpoint is a retrained reference build. It is close to the paper but not exact. Benchmarks score the paper build. Downloads run the retrained one.

Relative or metric: what each call returns

Depth Pro returns meters with absolute scale, plus an estimated focal length in pixels. The focal estimate is a real output, not a bonus. The paper scores it the most accurate method on all six of its focal test sets. On portrait photos from PPR10K, 64.6 percent of its focal guesses land within 25 percent of the truth. The next best method, SPEC, reaches 34.6 percent. Your photos can carry no EXIF data at all. Depth Pro still produces metric depth from pixels alone.

The DA3 metric tier works the other way around. It predicts canonical metric depth. You turn it into meters with one multiplication. The repository gives the formula: metric depth equals focal length in pixels times the network output, divided by 300. You supply the focal length. Depth Pro estimates it. That one difference decides many pipelines. Photos with no lens data favor Depth Pro. Cameras with known optics favor the simpler math.

One more scope difference. DA3 reads any number of views. It can return camera poses and fused geometry. Depth Pro reads one image and returns one depth map. It has no pose output and no multi-view mode. The full multi-view story, against VGGT, sits on the DA3 vs VGGT page.

Sharpness: the edge Depth Pro keeps

Boundary sharpness is Depth Pro's signature claim, and it survives DA3. Its paper scores boundary quality as F1 on depth edges. Depth Pro prints 0.409 on Sintel there. Depth Anything V2 prints 0.228, DPT 0.181. The DA3 paper does not score boundary F1 at all. No shared number exists between these two. That cell stays empty on purpose.

The structural reason is resolution. Depth Pro predicts at a fixed 1536 by 1536, about 2.25 megapixels. Its paper shows boundary accuracy climbing about threefold per resolution doubling. DA3 trains at a base resolution of 504 pixels, with varying aspect ratios. Higher native output is where fine strands, hair, and fence lines survive.

DA3 answers with training, not resolution. Its metric tier trains on teacher-model labels. In the paper's own figures, those labels add detail and sharper structures to metric maps. The paper calls this real but shows it in pictures, not as a boundary score against Depth Pro. If edges decide your product, test both models on your own scenes. No published table settles it.

Depth Pro also has stated limits. Its conclusion names translucent surfaces and volumetric scattering as weak spots. Per-pixel depth is ill-posed there. Fog, glass, and smoke are not its friends.

Speed and size, each on its own rig

One basis note on the parameter column: the DA3-Small cell carries the repository card's form, 0.08B. The paper's own size table prints that tier at 0.03B. Different forms, same model. The full two-form argument lives on the DA3 vs VGGT page.

The dash is deliberate. The DA3 paper prints speed rows for the main series tiers only. The metric tier never gets its own row. Nobody has published a DA3-metric latency figure.

Two readings matter. Depth Pro runs at a constant 341.3 milliseconds per image. Feed it VGA, HD, or 4K, the time holds, because the fixed 1536 by 1536 pipeline resizes everything first. That predictability is a feature for planning. The DA3-Small row is context, not a rival. It is the small relative tier, on a different GPU, and it is the tier this platform serves. Never divide speeds across rigs. The two papers used different cards.

The license split that decides commercial use

This is where the two models part ways hardest. Three documents, all read 28 October 2026.

The weights row is the one that bites. The license on the Hub repository grants use "exclusively for Research Purposes". The same text says research use "does not include any commercial exploitation, product development or use in any commercial product or service". Shipping a product on Depth Pro weights needs a separate arrangement with Apple. The code repository carries the milder sample-code license. But the weights are the model.

DA3 splits by tier instead. The metric tier, the monocular tier, and the small and base main-series tiers are Apache 2.0. The large and giant main-series tiers carry a non-commercial license. The full tier map, with its one open conflict, is on the DA3 vs DA2 page. For a commercial metric pipeline, the DA3 metric tier is the clean row. Depth Pro is the one that needs a lawyer first.

The hosting market, counted

Apple sells no Depth Pro API. The company ships code and weights. It sells no calls. We checked for any first-party paid service on 28 October 2026 and found none.

Third parties fill some of the gap, all on Replicate, all billing GPU time rather than images. The chenxwh/ml-depth-pro lane bills 0.000975 per second on an L40S. Its median run sits near 0.0034, read the same day. A second community lane, ibrahimpenekli/depth-pro, bills the same rate. Its median run sits near 0.00098. No seller anywhere prices Depth Pro per image. We ran that query and it came back empty. On fal, the Depth Pro page address returns a not-found answer.

For DA3, one seller prices per image. WaveSpeed runs the DA3 image endpoint at a fixed 0.005 an image, re-read 28 October 2026. Community lanes for DA3 variants also run on Replicate. The asymmetry is real. Depth Pro time is sold by the second. DA3 depth is sold by the image.

The download counts tell the same story from the demand side, all read 28 October 2026. The DA3 metric tier sits at 445,186 downloads last month. The transformers-ready Depth Pro packaging, apple/DepthPro-hf, sits at 51,586. The original apple/DepthPro sits at 10,741. One irony closes the census. Apple publishes Core ML builds of Depth Anything small and v2-small for its own devices. It ships no Core ML build of Depth Pro.

The lane this platform serves

This platform serves one depth row: da3-small, the small tier of the DA3 main series, on the vision route. It has been on the public list since 6 October 2026. It is metered per image against a prepaid wallet. The lane gives you what the local small tier gives you. That means relative depth, confidence, and camera parameters. It does not give metric depth. For meters, run the DA3 metric tier yourself under Apache 2.0. Or call a Depth Pro lane on Replicate and accept its terms.

Treat the lane as a young shelf and test your own scenes. A first call is three fields and a file:

curl -sS $IRONSTRATUM_BASE_URL/v1/vision \
  -H "Authorization: Bearer $IRONSTRATUM_API_KEY" \
  -F "model=da3-small" \
  -F "file=@frame.jpg" \
  -F "response_format=json"

Billing counts images, not requests. A request refused before it is served bills nothing. The current rate renders on the pricing page, drawn from the platform's model data:

Create a key on the signup page. The tier's full sheet is on the da3-small model page. For the budget math, image by image, read DA3 small API pricing. For running the small tier on your own CPU, read DA3 small, the CPU tier. For pairing depth with masks on the same route, read robot vision, segmentation and depth.

How to choose

Decision rules, in the order that tends to decide real projects.

• Ship a commercial product that needs meters: run the DA3 metric tier. It is Apache 2.0, it leads the one shared table, and its conversion to meters is one line. Depth Pro weights need Apple's permission for business use.
• Photos with no lens data, no EXIF, unknown cameras: Depth Pro. It guesses the focal length better than anything else published. Its metric output needs no intrinsics.
• Compositing, hair, fur, fence lines, thin strands: Depth Pro. Its 2.25-megapixel native output and its boundary record are its strongest case. Test DA3-metric too. No published table scores that fight.
• Robot or drone with known cameras, or any multi-camera rig: DA3. Known optics make the focal conversion trivial. Only DA3 adds poses and multi-view fusion. The robotics walkthroughs live on the DA3 vs VGGT page.
• Relative depth only, order without meters: the DA3 main series. On this platform, that is the da3-small lane. It is cheaper to run, with faster tiers and Apache terms at the small and base sizes.
• Fog, glass, smoke, and translucent things in general: neither model is strong. Depth Pro names the limit itself. Plan a fallback for those scenes.

Frequently asked questions

Is Depth Anything 3 better than Depth Pro?

On metric accuracy, the one shared table says yes. The DA3 paper ran both models on five test sets. Its metric tier won all five, read 28 October 2026. On boundary sharpness and focal estimation, Depth Pro still leads. No shared table scores either of those. The honest answer names the job first.

Is Depth Pro free for commercial use?

No. The weights carry Apple's research-only license. It excludes any commercial exploitation, product development, or use in any commercial product or service. The code repository carries a milder sample-code license. But the weights are what you ship. The DA3 metric tier, by contrast, is Apache 2.0.

Is there a Depth Pro API?

Not from Apple. The company publishes code and weights. It sells no calls. Community lanes on Replicate bill per GPU second at 0.000975, read 28 October 2026. No seller prices Depth Pro per image.

Does Depth Pro or Depth Anything 3 output meters?

Depth Pro outputs meters directly, with absolute scale. The DA3 main series outputs relative depth, an ordering from near to far. The DA3 metric tier outputs canonical metric depth. Multiply by the focal length in pixels, divide by 300, and you have meters.

Which one is faster?

Each paper measured its own rig, so no clean ratio exists. Depth Pro reports 341.3 milliseconds per image on a V100, flat at any input size. DA3 reports its small tier at 160.5 images per second on an A100. That is the relative tier this platform serves, not the metric tier. The DA3 metric tier has no published speed figure.

Which should I use for robotics?

DA3, in most cases. Known cameras make its metric conversion one line. It adds camera poses and multi-view geometry, which Depth Pro lacks. Its ecosystem has two ROS 2 wrappers, listed in the repository. Depth Pro stays relevant for single-camera metric work where edges matter. Close-range manipulation over textured scenes is the classic case.

What is Depth Pro, in one sentence?

Apple's October 2024 model for zero-shot metric depth. One image in, depth in meters out. It estimates the focal length from the image itself, in 0.3 seconds on a standard GPU.

Sources read in full on 2026-10-28: the Depth Pro paper, the DA3 paper, the Depth Pro repository, the Depth Pro Hub card and license, the DA3 repository, the DA3METRIC-LARGE card, the Replicate lane, and the platform's own model data.

Last verified: 2026-10-28