Depth Anything vs MiDaS: the classic baseline and its successor

Depth Anything passed MiDaS, and the field moved with it. MiDaS last shipped a model in December 2022. Its repository was archived in August 2024. One table scores both families under the same protocol. It is the V2 paper's own table. There, the V2 large and giant tiers beat MiDaS v3.1 on relative error across all five test sets. MiDaS still holds two edges. It runs on almost any hardware, from phones to CPU camera feeds. And its code is MIT licensed. This page gives the version history and the pick-by-job rules. It prints the numbers that compare cleanly, and flags the ones that do not.

Last verified: 2026-10-28

The short answer

Run Depth Anything for new work. The current generation, DA3, reads one image or many. It returns depth, confidence, and camera geometry in one pass. Its small tier runs at 160.5 images per second on the paper's test rig. The V2 generation loads in one line through the standard transformers pipeline. Four DA3 tiers, including the small one, carry Apache 2.0.

Keep MiDaS where MiDaS already works. Its 21-million-parameter small tier runs at 90 frames per second on one RTX 3090. An Intel CPU build sustains 22 frames per second at camera resolution. Sample apps exist for iOS, Android, and ROS 1. The code is MIT licensed, with no tier split. Nothing new is coming. The last model landed in December 2022. The repository was archived in August 2024.

Migrate when quality is the job. In the V2 paper's shared table, the relative-error gap on driving scenes is wide. It is 0.074 against 0.127. On the threshold metric the story is narrower than the blogs claim. Two datasets still favor the old model. Both facts sit in the benchmark section below, with the reason.

What MiDaS is

MiDaS is the monocular depth project from a team at Intel's Intelligent Systems Lab. It was first released in July 2019. The founding paper is Towards Robust Monocular Depth Estimation. The authors are Ranftl, Lasinger, Hafner, Schindler, and Koltun. It introduced two ideas. Each later model in this space inherited both.

The first idea is the training mix. Depth datasets disagree: different ranges, scales, and sensor types. MiDaS trained on many of them at once. It treated depth as affine-invariant. A prediction only had to be correct up to a scale and a shift. The final v3.1 release trained on 12 public datasets. They run from ReDWeb and MegaDepth to KITTI and NYU Depth V2.

The second idea is the benchmark. The paper's title is the giveaway: mixing datasets for zero-shot cross-dataset transfer. Score a model on datasets it never trained on. You then measure robustness instead of memorization. Each table in this matchup, on both sides, descends from that test.

The release line follows. It comes from the repository's own changelog.

The v3.0 release rode the Vision Transformers for Dense Prediction paper. That paper rebuilt the encoder as a DPT. The v3.1 release swapped in five modern backbones: BEiT, Swin2, Swin, Next-ViT, and LeViT. It is documented in a separate technical report from July 2023.

The install base is real and still moving. The most-used MiDaS card on the Hub is the v3.0 hybrid model. It logged 600,776 downloads in the last month, read 28 October 2026. That is more monthly traffic than any DA3 card. DA3-SMALL sits at 234,136. One packaging of the V2 small model, for contrast, runs at 1.53 million. Old tools die slowly in this field. MiDaS is not dying.

How the successor grew past the baseline

Depth Anything arrived in January 2024, five years after MiDaS. The team was different. The first release note is dated 22 January 2024 in the V1 repository. That repository's comparison section is pointed. It measures the new model against "the previously best MiDaS v3.1 BEiT_L-512 model". It then prints a fairness note the blogs never quote:

Please note that the latest MiDaS is also trained on KITTI and NYUv2, while we do not.

That one sentence explains most of the benchmark section below. Keep it in mind there.

The handover outside the papers was fast. On 23 January 2024, one day after release, the V1 team shipped a retrained depth ControlNet. They wrote that it "offers more precise synthesis than the previous MiDaS-based ControlNet". The same note records the integrations. They landed in the two main diffusion interfaces of the day. Within two weeks, Depth Anything was the default depth processor for two more image tools. The diffusion world had been running MiDaS depth for years. It switched in a month.

Then the releases stacked. V2 landed in June 2024. It was accepted at NeurIPS. DA3 landed in November 2025, marked ICLR 2026 Oral on its project page. It changed the job again. Any number of views now worked, with camera poses, confidence maps, and 3D exports. MiDaS, meanwhile, shipped nothing after December 2022.

The numbers, and the trap in comparing them

Three benchmark suites appear in this comparison. They do not mix. This section shows exactly why.

The MiDaS repository scores its own models on six datasets with error metrics. Lower is better. Two of its six, DIW and TUM, are never scored in the Depth Anything papers. Its delta1 column is an error rate, so smaller is better. Source: the repository's accuracy table, read 28 October 2026.

The Depth Anything V2 paper scores five datasets with two metrics. AbsRel is lower-better. Delta1 is an accuracy, so higher is better. Source: Table 2 of the paper, read the same day. This is the only table anywhere that runs both families under one protocol. The V2 authors ran MiDaS themselves.

On AbsRel, the error metric, the V2 large and giant tiers win all five datasets. The smaller tiers do not sweep. V2 Small loses NYU, 0.053 against 0.048, and ETH3D, 0.142 against 0.139. V2 Base loses NYU, 0.049 against 0.048. On the delta1 metric above, the large and giant rows win three: KITTI, Sintel, DIODE. MiDaS keeps NYU by a thousandth, 0.980 against 0.979. It keeps ETH3D by two thousandths in the large row. The fairness note matters here. MiDaS v3.1 trained on NYU Depth V2 and KITTI. So the NYU cell is a trained-set win. The ETH3D cell is not. And on KITTI, the other trained set, V2 wins by a wide margin anyway, 0.946 against 0.850. The V2 paper's own caption reads the table honestly. "Solely from the metrics, Depth Anything V2 is better than MiDaS, but merely comparable with V1."

Now the third suite. The DA3 paper re-scores V2 inside its own benchmark. It reports DA3 ahead on rank, 2.20 against 2.60, with V2 keeping NYU and Sintel. The word MiDaS appears zero times in that paper. This worker grepped the full text on 28 October 2026 to check. No shared-protocol DA3-versus-MiDaS number exists. This page prints none.

One more trap, smaller but real. The V2 paper and the DA3 paper both score "DA2" on these five datasets. Four cells match exactly. The Sintel cell does not: 75.2 in the V2 paper's table, 77.2 in the DA3 paper's. Close is not equal. No cell crosses between protocols on this page. The head-to-head table below leaves the accuracy row blank for exactly this reason.

The head to head

The blank row is deliberate. Anyone who prints one accuracy column with MiDaS and DA3 cells is mixing protocols. The numbers will not mean what the table claims.

Speed and hardware: the old tool's last strong suit

MiDaS still earns its keep on weak hardware. Its repository table is the census to cite. All speeds below are its own measurements. They ran on one RTX 3090, at each model's training resolution.

Below the GPU floor it goes further. The repository clocks the v2.1 small model at 22 frames per second at 640 by 480. The test rig is an 11th-generation laptop CPU, through Intel's OpenVINO build. TensorFlow and ONNX exports exist, though only for v2.1. Sample apps ship for iOS and Android. A ROS 1 package ships in the repository. The PyTorch Hub page still serves the original three-liner. It is torch.hub.load("intel-isl/MiDaS", "DPT_Large"), under the project's original org name.

The Depth Anything side answers with its own numbers, on its own rig. The DA3 paper measures its small tier at 160.5 images per second on one A100. The scenes hold 32 images. It holds about 4,000 views in memory. That is a different GPU and a different resolution. It is also a different unit than the MiDaS table. Dividing 160.5 by 90 produces a number that means nothing. This page refuses to print it.

For the CPU lane, the small DA3 tier has its own economics on this platform. The CPU tier walkthrough covers that story.

Licenses: read each surface separately

Every license fact below was read at its source on 28 October 2026.

The MiDaS repository is MIT, and the copyright line names Intel ISL, 2019. On the Hub, the two most-used MiDaS cards carry Apache 2.0 tags. The tiny Swin2 card carries MIT. For a business, all of these are safe. MIT and Apache 2.0 both allow commercial use, with no fee and no gate. This is the one axis where MiDaS is still the simpler purchase. Every weight file and every card clears legal in one read.

Depth Anything V2 splits by tier, in the repository's own words. The Small model is Apache 2.0. Base, Large, and Giant carry CC BY-NC 4.0, which bars commercial use. The 1.3B Giant row was announced but never shipped. DA3 widened the commercial set to four tiers. Small, base, the metric large, and the mono large are Apache 2.0. The large, giant, and nested giant tiers stay CC BY-NC 4.0. A refreshed 1.1 series fixes a training bug. The repository says to prefer it, above all for street scenes.

The practical rule is simple. If your product ships, list the exact tier and read its card on download day. The DA3 versus 2 page carries the full tier table. It also carries the one row where the publisher's own two surfaces disagree.

Where each one wins

Pick MiDaS when the hardware is the constraint. Pick it for a phone build, an old GPU, or a CPU-only robot companion computer. It fits a legacy ROS 1 stack too. The 21M small tier, the ONNX export, and the mobile samples were built for these. Pick it when a tool you cannot change already calls it. Pick it when you want one license answer instead of a tier map.

Pick Depth Anything for anything new. Single-image quality: V2 Large or Giant, or the DA3 tiers, beat MiDaS on each protocol that scores both. Metric depth in meters: V2 ships metric fine-tunes. DA3 ships an Apache metric tier. The Depth Pro comparison owns that matchup. Multi-view geometry, camera poses, 3D exports: DA3 only. MiDaS never had the interface. The VGGT comparison carries that lane. The rest of the rival field is MoGe, Marigold, and ZoeDepth. The round-up page is the single table for those.

Inside the Depth Anything family, the version choice has its own page. It is DA3 versus 2, including the output flip that breaks migrated code.

The maintenance reality

Facts from the GitHub record, read 28 October 2026. The MiDaS repository is archived. That makes it read-only. The last push was 23 August 2024. It added a security policy file. The last model change before that was the v3.1 readme update in July 2023. The repository holds 5,415 stars and 164 open issues.

None of that breaks existing code. The weights still download, and the license still stands. An archived repository is not a withdrawn one. What changes is the forward path. There will be no bug fixes, no new backbones, and no support for new hardware. If a future PyTorch release breaks the loader, nobody is on the other end.

Intel's own follow-ups went elsewhere. ZoeDepth added metric output on top of MiDaS in 2023. The LDM3D image generator was trained on MiDaS depth maps. The repository's applications section records both. The active frontier is refreshes and new tiers. That sits entirely on the Depth Anything side today.

Calling either one over an API

Neither publisher sells calls to its own model. Both ship code and weights. If you would rather call than host, the market is thin and lopsided.

For MiDaS, one community lane on Replicate lists about 0.00022 a run on a T4 GPU. Predictions typically land inside one second. Read 28 October 2026. A depth utility endpoint titled Midas Depth Estimation exists on fal. Same-day read. That page prints no rate. This is the whole list this page found. For the DA3 side, the hosted set is wider. It is censused on the DA3 pricing guide.

This platform carries one depth row. It is da3-small, the small tier of DA3, on the vision route, metered per image against a prepaid wallet. It has been on the public list since 6 October 2026. MiDaS is not served here. This page will not pretend otherwise. A first call is four fields and a file:

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

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

An account opened on the signup page covers the lane. The small tier also has its own model page with the full sheet. The alternatives page maps each hosted option if da3-small is not the right fit.

Frequently asked questions

What is MiDaS?

A monocular depth estimation project from a team around Intel's Intelligent Systems Lab. It was first released in July 2019. It reads one image and returns a relative depth map. It was trained on up to 12 public datasets with an affine-invariant objective. Its last release, v3.1, shipped in December 2022. The founding paper is arXiv 1907.01341.

Is Depth Anything better than MiDaS?

The one shared-protocol table is the V2 paper's. There, yes on relative error for the larger tiers, across all five datasets. It is three of five on the threshold metric for those tiers. MiDaS keeps NYU and ETH3D. NYU is a set it trained on; ETH3D is not. DA3 is stronger again on its own suite. No DA3-versus-MiDaS table exists. The DA3 paper never mentions MiDaS.

Is MiDaS still maintained?

No. The repository was archived and made read-only. The last push was 23 August 2024. There has been no model release since December 2022. Weights and code remain downloadable under MIT. The 164 open issues stay open.

Which one should I run on a CPU or a phone?

MiDaS v2.1 small, if the hardware cannot be changed. It is a 21M-parameter model with an OpenVINO CPU build measured at 22 frames per second. iOS and Android samples ship too. For CPU work with current quality, the DA3 small tier is the modern answer. Its CPU economics are worked through on the CPU tier page.

What license is MiDaS under?

The GitHub repository is MIT, copyright Intel ISL 2019. The most-used Hub cards carry Apache 2.0 tags. All are commercial-safe, with no gate. Depth Anything splits by tier instead. V2 has one Apache tier, DA3 has four, and the rest bar commercial use.

Can I call MiDaS or Depth Anything over an API?

Neither publisher runs a paid API. Third parties do. A community MiDaS lane on Replicate charges about 0.00022 a run, read 28 October 2026. A fal depth utility endpoint exists too. This platform serves the da3-small tier per image on the vision route. It has been on the public list since 6 October 2026.

Sources read in full on 2026-10-28. The MiDaS repository and its LICENSE. The MiDaS paper. The DPT paper. The MiDaS v3.1 technical report. The Depth Anything V1 repository. The Depth Anything V2 paper and its repository. The DA3 paper and its repository. The PyTorch Hub MiDaS page. And the publisher's model card records.

Last verified: 2026-10-28