Alternatives to Depth Anything: the decision map

If you are weighing Depth Anything against the field, the answer splits by the job. For sharp relative depth from one image, the Depth Anything tiers are still the default. For metric depth in real units, rivals carry the job under cleaner terms. For multi-view geometry and camera pose, VGGT and MapAnything now compete with Depth Anything 3 itself. For depth bought as a hosted per-image call, the shelf is thin. This page maps every row on that shelf.

Last verified: 2026-10-28

That is the whole decision in six sentences. The rest is evidence. It covers what each rival does well, where the licenses bite, which sellers bill per image, and what this platform serves.

What you are actually replacing

Depth Anything names two products, not one. Version 2 shipped in June 2024 as a single-image relative depth family. Its tiers run from Small at 24.8 million parameters to Large at 335.3 million. A 1.3 billion Giant tier was announced and never released. Depth Anything 3 shipped in November 2025. It moved the family to multi-view geometry. It reads depth from any number of views, with or without camera poses. It also outputs camera pose and 3D scene exports. Its tiers run from SMALL at 0.08 billion parameters to NESTED-GIANT at 1.40 billion.

Neither version ships a hosted API. The project pages link weights, code, and community tools such as ComfyUI nodes and ROS 2 wrappers. Every paid depth route in this guide is a third party or this platform. The version-by-version story, with the paper's own delta table, sits on Depth Anything 3 vs 2.

The decision map, job by job

Sharp relative depth from one image. Stay in the family. An Apache small tier is the default here, run yourself or bought per image. The classic rival, MiDaS, moved to archive status in August 2025. Its last release dates to December 2022. Marigold is the taste option. It refines depth through a diffusion process. Its default settings favor speed, and one pass can run fast. Turn the ensemble up for quality and the cost climbs toward seconds per image. The Depth Anything tiers stay faster at their defaults. If the deliverable is a depth-looking map for video or art, Marigold is in the running. If it is a number feeding a later step, it is not.

Metric depth in real units. Leave relative maps behind. Inside the family, the DA3 METRIC-LARGE tier holds Apache 2.0 terms and outputs metric depth. Outside it, Metric3D carries a BSD-2-Clause license across two checkpoint generations. Depth Pro, from Apple, sells the zero-shot metric story. Its code license is a sample-code license. Its weights license bars most product use. Read both before you ship. ZoeDepth held this job for years under MIT terms, then stopped. Its repo went to archive status in May 2025. Today the clean open rival is Metric3D. The Apache answer inside the family is the DA3 metric tier. The full rivalry sits on Depth Anything 3 vs Depth Pro.

Multi-view geometry, pose, and 3D work. This is the job Depth Anything 3 was built for. The field is crowded. VGGT won a CVPR 2025 Best Paper award for feed-forward geometry from unposed views. It ships a separate checkpoint for commercial use beside its research license. MapAnything, from Meta and CMU, maps many views. Its research release carries an Apache code license with weights that bar commercial use. A separate release ships the same weights under Apache 2.0 terms. The DA3 paper reports beating the prior state of the art by 44.3 percent on camera pose accuracy and 25.1 percent on geometric accuracy, on average. The mono matchup against VGGT sits on Depth Anything 3 vs VGGT. Pick by license first. Then pick by pose or depth emphasis.

Video depth. The family covers it. Video Depth Anything extends the line to clips. The real cost of running video batches yourself is memory. The crash family that comes with it is documented on Depth Anything video out of memory errors. Hosted per-video depth is a thin market. fal sells a Depth Anything video endpoint at a printed rate of 0.04 per second of video, read 28 October 2026. The other depth-named video seller this research surfaced, depthify.ai, has paused operations, read on its own page the same day.

Masks, not depth. If the real job is segmentation, depth is the wrong axis. Compare Segment Anything 2.1 instead. That matchup sits on alternatives to SAM 2. The two-model pattern, depth plus masks, is drawn on robot vision, segmentation and depth.

Depth as a paid call, no GPU. That job is the hosted shelf two sections down. Read the license table first. What you may ship decides which tiers can enter your shortlist.

Where Depth Anything wins

• Open weights at every useful size, from 24.8 million to 1.40 billion parameters.
• Four Apache 2.0 tiers in the current generation, metric included. A product can ride the family with no license talk.
• The multi-view jump: one line covers mono depth, multi-view depth, camera pose, and 3D exports.
• Momentum: the DA3 small tier counts 234,136 downloads in the last month on Hugging Face, read 28 October 2026. The tooling list already spans ComfyUI, ROS 2, and 3D tools.

Where it loses

• The big tiers are closed for products. Version 2 licenses only its Small tier under Apache terms. The current version licenses its LARGE, GIANT, and NESTED tiers under CC BY-NC 4.0, which bars commercial use.
• No official API has ever shipped, for either version. Buying depth means a third-party seller or a platform lane like this one.
• Metric depth has specialist rivals with clean licenses and long track records. Metric3D is first among them.
• Multi-view geometry has a rival that held the state of the art before November 2025 and still wins on license clarity for its commercial checkpoint.
• The archives tell a story. MiDaS and ZoeDepth both stopped. And the version 2 repo went offline for six days in June 2024 after GitHub flagged it. Any single lab can stop.

The license split, tier by tier

Licenses decide products before benchmarks do. This family splits by tier, not by name. Every row below was read on the named model's official page on 28 October 2026. The commercial-use column is our reading of the license text. It is not legal advice.

The split is not theory. Video editors on the Blackmagic forum have asked for Depth Anything inside Resolve. The same threads record that commercial depth tools ship the V2 Small tier. It is the only version 2 tier under Apache terms, and the thread says that is why those tools run it. A product decision hung on one tier's license, just as the table predicts. For robot-model licenses beyond depth, see the GR00T license walkthrough.

Who sells depth per image

The search results for depth API alternatives are directory pages and repo-matching lists. They are not markets. This census reads each seller's own page. All figures in US dollars as published by each named seller, read 28 October 2026.

Three relations fall out of the table. WaveSpeed's fixed DA3 rate reads below ModelRunner's V2 rate. At these two sellers, the newer version is the cheaper buy. The community metric tier on Replicate bills seconds, not images, and fal's video endpoint bills seconds of video. Their costs move with run time, and a slow batch is a slow bill. And two names a buyer would check, DeepInfra and Hugging Face's providers, carry no Depth Anything row at all as of this read. The wider question of which models the big hosts serve walks on the big-hosts survey. The seller-by-seller DA3 census, with the earlier reads, sits on who serves DA3 today.

What this platform serves

This platform serves the DA3 small tier as a metered lane under the da3-small model id, on the vision route. It has been on the public list since 6 October 2026. The full model sheet sits on the da3-small page.

Billing counts images, not requests. The route takes multipart form data with the file attached. It holds a 25 MiB payload ceiling. It returns the model's output plus a usage block with the billed image count. 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=@frame.jpg" \
  -F "response_format=json"

The rate renders on the pricing page, drawn from the platform's model data:

Open an account on the signup page. Create a key. Send one image before you batch. For the budget math image by image, see the DA3 small pricing walkthrough. The same route serves the SAM 2.1 base mask model per image. The SAM 2.1 large row sits on the rate list without serving traffic as of 28 October 2026. Mask work has a lane here too.

How to choose in 60 seconds

1. One image, relative depth: run an Apache small tier yourself, or buy it per image only when there is no GPU in the plan.
2. Real units from one image: Metric3D for the clean open license, the DA3 metric tier for the Apache answer, and a license read before any Depth Pro shipment.
3. Many views or camera pose: DA3 against VGGT, decided on license first and pose-versus-depth emphasis second.
4. A product to sell: the license column above decides before any benchmark does. It already decided some of the tools you use.

Questions builders ask

Is Depth Anything free for commercial use?

Only some tiers. In version 2, the Small tier alone carries Apache 2.0 terms. In the current version, four tiers carry Apache 2.0: SMALL, BASE, METRIC-LARGE, and MONO-LARGE. The LARGE, GIANT, and NESTED tiers carry CC BY-NC 4.0. That license bars commercial use. Every read above is from the official repos, dated 28 October 2026.

What is the best alternative to Depth Anything V2?

There is no single best, only a best per job. Sharp single-image maps stay in the family. The small tiers hold the accuracy-per-second lead. Metric work moves to Metric3D or the DA3 metric tier. Multi-view work moves to VGGT or MapAnything. The per-rival stories, benchmarks included, are collected in the depth rivals round-up.

Is there an official Depth Anything API?

No. Neither version has shipped one. The project pages link weights and code, not endpoints. The hosted rows that exist are third parties. ModelRunner and WaveSpeed publish per-image rates. A community metric tier rides Replicate, billed by the second, and fal sells image and video endpoints. Roboflow serves V2 through its own deployment paths and prints no per-image rate. This platform's da3-small lane is a served per-image row on the vision route.

Can I use Depth Anything V2 Base or Large in a commercial product?

Not under the published terms. Both tiers carry CC-BY-NC-4.0, a non-commercial license. V2 Small is the tier that ships in products. Commercial depth tools already run it for that reason. The current version gives products a wider Apache set, metric included.

What replaces ZoeDepth or MiDaS in 2026?

Nothing maintains them. ZoeDepth went to archive status in May 2025. MiDaS followed in August 2025, with its last release in December 2022. Their jobs moved forward. Metric depth now rides Metric3D, Depth Pro, or the DA3 metric tier. The indoor-benchmark lineage they owned is folded into the current suites. The MiDaS matchup in depth sits on Depth Anything vs MiDaS.

How do hosted depth API rates compare?

Among the per-image rates published on 28 October 2026, WaveSpeed's DA3 read at 0.005 per image is the lowest. ModelRunner's V2 read at 0.01 per image comes next. The community metric tier on Replicate can land near either, because it bills seconds. This platform's own row renders in the live table above. The honest comparison is cost per usable output, not cost per call. Tiers differ in what they return: relative maps, metric maps, or pose plus depth.

Last verified: 2026-10-28