Depth Anything vs MoGe, Marigold, and ZoeDepth: the four-way depth pick

Depth Anything, MoGe, Marigold, and ZoeDepth answer four different depth questions, and the pick follows the question. Depth Anything is the breadth family, fastest at the small tier. MoGe returns a 3D point map, camera field of view, and normals in one pass. Marigold is a repurposed image generator, fine detail at boundaries, slowest per image. ZoeDepth shipped metric depth in 2023 and now sits unmaintained.

This page is the round-up for the three strings people actually type: depth anything vs moge, depth anything vs marigold, and depth anything vs zoedepth. Each rival gets one tight section. The per-pair deep dives live on their own pages, linked through the family map below. Every number here names its source and the date it was read, 28 October 2026.

Last verified: 2026-10-28

The verdict, fast

Pick by the job, not by the brand.

• One depth map per image, fast, with the largest ecosystem of ports and wrappers: a Depth Anything small tier. It is the row this platform serves, and it is the speed row in its own paper's table.
• A full 3D scene from one photo: points, depth, normals, and camera field of view in a single forward pass: MoGe. Its 2025 paper also prints the only table that ranks three of these four families on one protocol.
• Fine detail at boundaries, when seconds per image are acceptable: Marigold. It is a fine-tuned image generator, and it works like one.
• Metric depth, in meters, from a single model without calibration: MoGe-2 today, or a Depth Anything metric tier. ZoeDepth was the 2023 pioneer of that job and its repository now asks you to fork it.

One caution before any table: these four papers do not share a scoring protocol. The comparison table below comes from one paper, scored one way. The cross-paper trap has its own section.

Meet the four rivals

Depth Anything: the breadth family

A ByteDance Seed team released Depth Anything 3 in November 2025. The paper is dated 13 November 2025. The family runs from a phone-class small tier to a giant tier. Metric and monocular variants sit beside the main series. Four tiers ship under Apache 2.0. The license split, the version-to-version changes, and the migration traps are covered in full on the Depth Anything 3 vs 2 page. On this page, Depth Anything is the incumbent. It is the model your stack probably already speaks.

The small tier is what this platform serves, under the model id da3-small. In the paper's speed table it runs 160.5 frames per second on one A100 across 32-image scenes. That is the fastest row of any family compared here.

MoGe: one pass, the whole geometry

Microsoft Research released MoGe in October 2024. It appeared at CVPR 2025. The idea is direct. The model predicts a 3D point map of the scene from one image. An affine-invariant form sidesteps scale ambiguity during training. Camera field of view falls out of the same pass. The repo reports 60 milliseconds per image for the ViT-L size at FP16, on an A100 or an RTX 3090.

The line is now three generations deep. MoGe-2, paper dated 3 July 2025, adds metric scale and sharper detail. MoGe-3, paper dated 20 July 2026, was released 18 August 2026 per the repo news line. It refines fine detail with a sparse volumetric step in 3D. Weights sit on the Hub under MIT-tagged cards. One practical wrinkle: MoGe-3 does not support macOS, because its custom matrix kernel builds on Triton.

Marigold: a picture generator that paints depth

An ETH Zurich team released Marigold in December 2023. It appeared at CVPR 2024 as an oral and best-paper candidate. The method fine-tunes Stable Diffusion, an image generator, until it paints depth instead of pictures. Training used about 74 thousand synthetic frames. It took roughly 2.5 days on one RTX 4090, per the paper.

Inference is where the design shows. The paper sets sampling at 50 DDIM denoising steps. The final map pools 10 runs made with different starting noise. That is diffusion work, not a single forward pass. A latent consistency model checkpoint trades some of that cost for speed. The repo ships it beside the standard one.

The line moved twice since. In May 2025 the same team extended the recipe to surface normals and intrinsic decomposition. In September 2026 a partly new team published Marigold V2, accepted at SIGGRAPH Asia 2026. It rebuilds the recipe on diffusion transformers rather than the older UNet.

ZoeDepth: the 2023 metric pioneer, now parked

Intel's Intelligent Systems Lab released ZoeDepth in February 2023. Its idea: take a relative-depth model pre-trained on 12 datasets, the MiDaS line, and add metric bins. The bins are a lightweight head that sorts the scene into depth ranges and places them in meters. A latent router picks the indoor or outdoor head per image. The flagship ZoeD-M12-NK model fine-tuned on NYU Depth v2 and KITTI jointly. That was a first at the time.

The repo banner now reads, in its own words, PROJECT NOT UNDER ACTIVE MANAGEMENT. Intel has stopped work and accepts no patches. The project still loads. It joined the transformers library in July 2024. The Intel Hub card has passed 22 million all-time downloads. The code is a finished artifact, not a live one. Its environment pins PyTorch 1.10.1, a late-2021 build.

One table, three rivals, one protocol

Only one paper among these puts three of the four families on the same scoring: the MoGe paper's depth table, read here on 28 October 2026. Scores are delta1 accuracy in percent, higher is better, under scale-invariant alignment. The rank averages eight test sets.

Marigold sits in a second block of the same table, scored under affine-invariant alignment instead. It reads 97.3 on NYU and 93.8 on KITTI, average rank 2.25, behind MoGe at 1.00. The two blocks are not one ranking. The paper separates them because the alignment differs.

Read the table honestly. ZoeDepth lands sixth of eight, older than every row above it. Depth Anything V2 lands fourth on depth rank, mid-field. Point-map models and metric models take the top of this scoring. The newer Depth Anything 3 is not in this table at all. Its own paper, under its own mono protocol, ranks DA3 at 2.20 against DA2 at 2.60 and VGGT at 3.75.

The cross-paper trap

Search results for these matchups print mixed tables. ZoeDepth's paper reports 0.953 delta1 on NYU as a fraction, under metric scoring on the trained domain. Marigold's paper reports 96.4 on NYU with a 10-run ensemble, under affine-invariant scoring. The MoGe table above reports yet another basis. None of those three numbers competes with the others. A page that stacks them in one column without naming protocols has made up a comparison.

The same caution applies to speed. A latency figure without its rig, precision, and model size is a decoration. This page names the rig wherever it prints a speed. One A100 for the MoGe and Depth Anything figures. One RTX 4090 for Marigold's training time. One RTX 3090 or A100 for MoGe's 60 milliseconds.

Licenses: three answers for four families

Licenses decide commercial projects before benchmarks do. Here is the split, each entry read at source on 28 October 2026.

The Marigold row carries the one real trap. The repo code is Apache 2.0. But the current depth checkpoint card carries an openrail++ tag, a use-restriction license inherited from the underlying generator. The fast LCM checkpoint on the same Hub account carries Apache 2.0. Two checkpoints, one repo, two weight licenses. Read the card tag on the day you download, not the repo badge.

The Depth Anything split is tier by tier. It has its own full treatment on the vs 2 page. The short form: small, base, metric, and mono tiers are Apache 2.0. Large and giant tiers carry a non-commercial license.

Where each rival wins

• Metric depth in meters, maintained model: MoGe-2, or a Depth Anything metric tier. Both are current, both permissively licensed at that tier.
• Indoor or outdoor metric depth on a fixed older domain, and the stack already runs transformers: ZoeDepth still delivers. Accept the parked repository and plan to own your fork.
• Boundary and thin-structure detail for creative work, depth for image editing pipelines: Marigold. Pay in denoising time, or take the LCM checkpoint.
• Full geometry, points and normals and field of view, one image, one pass: MoGe, any generation.
• Camera-rate depth for robots, widest port and wrapper ecosystem: a Depth Anything small tier. The multi-view and pose story against VGGT is a different question, covered on the VGGT page.
• Relative depth against the classic 2021 baseline: the MiDaS comparison owns that history, and ZoeDepth is its direct descendant.

What speed means for each

Marigold is the slow one by construction. Ten runs of 50 denoising steps each feed one final map. The paper's own ablation finds accuracy flattening near 10 steps, and its 50-step budget sits comfortably above that. The LCM checkpoint cuts the run count for less exacting work.

MoGe and Depth Anything are single-pass networks. MoGe quotes 60 milliseconds per image at the ViT-L size, FP16, on an A100 or an RTX 3090. The Depth Anything small tier quotes 160.5 frames per second on an A100 in its paper's table. Different rigs, so do not divide one by the other. The measured CPU numbers for the Depth Anything small tier, memory included, live on the CPU tier page.

ZoeDepth is also single-pass, on a 2023 design with 2022 dependencies. Age is the speed story there. It runs. It is small by modern standards. Nothing about it will get faster.

Who sells them as an API

No publisher among the four sells calls to its own model. All of them ship code and weights. The hosted market around these rivals is thin and lopsided. All rates below are in US dollars, published by each seller, read 28 October 2026.

ZoeDepth has a community lane on Replicate. The page quotes about 0.00044 US dollars per run, 2,272 runs per dollar, on a T4 GPU, with runs typically done in 2 seconds. Marigold has one too, from another publisher: about 0.078 US dollars per run, 12 runs per dollar, on an L40S GPU, with runs near 81 seconds. Same marketplace, roughly 177 times apart per run. The gap is the diffusion design showing up in a bill.

MoGe has no hosted lane that these searches could find. The queries are named in the money receipts file beside this draft. The cell stays dashed rather than guessed. Self-host the MIT weights, or use the free demo Spaces the team keeps.

Depth Anything 3 has exactly one per-image seller found this day: WaveSpeed, at a fixed 0.005 US dollars per image. The full seller census, per-second lanes included, lives on the provider check page.

This platform serves the small tier, model id da3-small, on the vision route, metered per image. It has been on the public list since 6 October 2026. 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"

Billing counts images, not requests. The rate renders on the pricing page, drawn from the platform's model data:

Open an account on the signup page, create a key, and send one image before you batch. The model's full sheet is on the da3-small page. For the budget math image by image, see the pricing walkthrough. If depth is one half of a perception stack, robot vision, segmentation and depth covers the two-model pattern.

The rest of the depth shelf

This page is the hub of the depth comparisons. Every per-rival story has its own page:

• Depth Anything 3 vs 2: what changed between generations, the paper's own delta table, migration traps.
• Depth Anything 3 vs VGGT: mono against multi-view, pose accuracy tier by tier.
• Depth Anything 3 vs Depth Pro: the metric-depth rivalry with Apple's entry.
• Depth Anything vs MiDaS: modern against the classic baseline everyone still runs, the line ZoeDepth builds on.
• Depth Anything 3 small on a CPU: measured CPU latencies and memory.
• Alternatives to Depth Anything: the whole field as user options.
• Depth Anything 3 small API pricing and the provider check: cost math and the seller census.

Frequently asked questions

Is Depth Anything better than MoGe, Marigold, and ZoeDepth?

On breadth, ports, and speed at the small tier, yes. On the one paper that scores three of the four on a shared protocol, MoGe ranks first. Depth Anything V2 lands fourth of eight. ZoeDepth lands sixth of eight for depth accuracy. Marigold is scored separately under a different alignment and sits second in its block. None of that settles your scene. Test your own images.

Which of these models outputs metric depth?

MoGe-2 outputs a metric point map and metric depth. ZoeDepth outputs metric depth within the domains it was tuned for, indoor NYU-style and outdoor KITTI-style. Depth Anything ships a dedicated metric tier. The standard Marigold and MoGe-1 checkpoints, and the Depth Anything main series, output relative depth. That orders a scene without measuring meters.

Is ZoeDepth still maintained?

No. Its repo carries a banner stating the project is not under active management. Intel has stopped work, and patches are not accepted. The code still loads through the transformers library. The weights remain downloadable under MIT. Treat it as a finished 2023 artifact.

Why is Marigold slow, and can it be faster?

Its final map pools 10 inference runs of 50 DDIM denoising steps each. That is the cost of the diffusion route. The latent consistency model checkpoint cuts both the step count and the ensemble need, trading some accuracy. Marigold V2, published September 2026, moves the recipe to diffusion transformers and reports sharper output.

Is there an API for MoGe, Marigold, or ZoeDepth?

Not from their publishers. Community lanes exist on Replicate for ZoeDepth, about 0.00044 US dollars a run, and for Marigold, about 0.078 US dollars a run. Both were read 28 October 2026. No hosted MoGe lane was found by the searches logged beside this draft. The teams also run free browser demos.

Which licenses allow commercial use?

MoGe code and weight cards: MIT. ZoeDepth repo and card: MIT. Depth Anything: the small, base, metric, and mono tiers are Apache 2.0, the large and giant tiers are not. Marigold code is Apache 2.0, but its current depth checkpoint carries an openrail++ use-restriction tag. Its LCM checkpoint is Apache 2.0. Check the card tag on download day.

Which one should I use on a robot?

A Depth Anything small tier for per-frame depth. Speed, memory headroom, Apache terms, and the wrapper ecosystem argue for it. MoGe where one photo should yield a full local 3D scene with camera field of view. Metric tiers where grasping or navigation needs meters. The robot framing, with the vision route contract, is laid out on the segmentation and depth page linked above.

Sources read in full on 2026-10-28. Depth papers: MoGe, MoGe-2, MoGe-3, Marigold, the Marigold adaptation paper, Marigold V2, ZoeDepth, and DA3. Code: MoGe, Marigold, ZoeDepth. The Hub cards for ZoeDepth and Marigold. The transformers ZoeDepth doc. Seller pages are named in the hosting section.

Last verified: 2026-10-28