Depth Anything 3 vs VGGT: which depth model to pick

Depth Anything 3 and VGGT are both feed-forward geometry models. The winner depends on the tier you run. The DA3 giant tier beats VGGT on four of five pose benchmarks in the DA3 paper's tables and trails by half a point on the fifth, read 28 October 2026. The DA3 small tier loses to VGGT on pose. It runs about 4.7 times faster. Licenses split both families. The per-tier numbers and the decision rules are below.

What each model is

VGGT came first. A team from Oxford's Visual Geometry Group and Meta AI released it in March 2025. The paper is dated 14 March 2025 on arXiv. It won a Best Paper Award at CVPR 2025. One network reads one view, a few, or hundreds. It returns camera parameters, depth maps, point maps, and 3D point tracks in one forward pass. The team ships one base checkpoint, VGGT-1B. The paper says the model has about 1.2 billion parameters in total. The Hub record for the checkpoint counts 1,256,537,516 parameters in the weights file, read 28 October 2026. Three forms, one model. Ask any seller which form its claim uses.

Depth Anything 3, or DA3, is the younger one. A ByteDance Seed team released it in November 2025. Its paper is dated 13 November 2025. It was accepted as an ICLR 2026 Oral. The family is not one model. The official repository lists a main series in four sizes, small through giant. It adds a metric series, a monocular series, and a nested giant that combines two of them. That tier structure is the heart of this comparison. The family-level claims and the tier-level tables tell different stories.

The two teams even share a backbone ancestor. VGGT patchifies images with DINO tokens. DA3 uses a plain DINOv2 encoder, unmodified. The differences sit in what each model predicts. They also sit in how the views talk to each other. Both points are covered below.

The head-to-head, from one paper's tables

The DA3 paper carries the only published table that runs both models on the same benchmark suite with the same scoring. These are the DA3 team's own numbers, on the suite they built. VGGT's team has not published a rebuttal table. Read them as one lab's scoring of both models. Dated 28 October 2026, from the paper's Tables 2 and 3.

Camera pose accuracy, scored as AUC30. Higher is better:

The pattern is clean. The giant tier beats VGGT on four of five sets. It trails by half a point on the fifth. The large tier trades wins with VGGT. The repository itself says the large tier scores comparable to VGGT. The small tier loses to VGGT on all five. Anyone quoting one family-level winner, with no tier named, quotes the wrong row.

Reconstruction quality. Scored as F-Score, no given poses. Higher is better:

Same pattern, wider gaps. On dense multi-view geometry, the giant tier dominates. The small tier falls far behind VGGT. One gap is the paper's, not this table's. The paper scores DTU with a different metric, chamfer distance. So DTU cannot sit in this F-Score table.

Then the comparison flips. Depth from one image, the monocular setting, is DA3's home ground. The paper's Table 4 ranks methods on five standard monocular benchmarks by average rank. DA3 ranks 2.20. Depth Anything 2 ranks 2.60. VGGT ranks 3.75, the worst of the three. On SINTEL, DA3 scores 75.5 on the accuracy metric. VGGT scores 67.9. If your work is single images, the depth family wins. The multi-view specialist comes last.

Two forms of the headline number

Here is a trap worth ten minutes of your time. The DA3 record states its lead over VGGT in two forms. Both are official. Both were read on 28 October 2026.

The arXiv abstract page says DA3 surpasses VGGT by an average of 44.3 percent in camera pose accuracy. It also says 25.1 percent in geometric accuracy. The paper's own full text, in the arXiv HTML rendering, says 35.7 percent and 23.6 percent. The project page says 35.7 and 23.6 too. Only one version of the paper exists on arXiv. It was submitted 13 November 2025. So this is not a version drift. Two official surfaces of one paper disagree about the headline average.

The honest reading: skip the headline average. Use the per-dataset tables above. Averages hide which tier ran. They hide which benchmark weighted the mean. The tables name both.

Speed, memory, and scale

The DA3 paper's Table 8 measures both models on one A100 GPU. The scenes hold 32 images. The resolution is 504 by 336 pixels:

The parameter column here is the paper's own Table 8 form. The repository cards count two rows differently. They list the giant tier at 1.15B and the small tier at 0.08B. The paper's sums are 1.20B and 0.03B. Different forms, same models. Two facts stand out. The small tier is about 4.7 times faster than VGGT on the same rig. That figure is 160.5 divided by 34.1. The small tier also holds roughly ten times more views in memory before the 80 GB card fills. For a robot or a batch pipeline, those two numbers can beat the pose gap.

VGGT's own speed story sits on its project page. It claims reconstruction in under one second. Its paper measured that on one H100 GPU at 336 by 518 pixels. Different rig, different resolution. Do not divide that figure against the A100 table. The VGGT team also posted a memory fix on 15 May 2026. It holds roughly two to three times more frames at the same memory budget. The DA3 side answered with a streaming variant, released 11 December 2025. It handles ultra-long video under 12 GB of GPU memory through a sliding window.

Two architectures, one shared backbone

The design argument between these papers is the real story of 2025-2026 geometry models. It is simple enough to state in two sentences each.

VGGT alternates between two attention arrangements. Frame-wise layers let each view digest alone. Global layers then let all views exchange signals. Camera tokens ride along. Separate heads predict cameras, depth, point maps, and tracks. DA3 keeps one plain DINOv2 transformer. It never touches the encoder's internals. Cross-view talk comes from rearranging the token layout between layers. The model predicts one thing per pixel: a depth value plus a viewing ray. Cameras fall out of the ray map through classical geometry, helped by a small head.

The DA3 paper tested the difference directly. At a similar size, a VGGT-style two-transformer stack scored 79.8 percent of the DA3 baseline in their ablation. The authors blame the gap on DA3's fully pretrained backbone. VGGT's second transformer is only partly trained. Your takeaway is practical, not partisan. DA3 outputs depth, rays, camera parameters, and confidence. An extra head adds 3D Gaussians for novel view synthesis. VGGT outputs add point tracks, which DA3 does not predict. Its frozen features also feed a real ecosystem of downstream work. Robot control policies built on VGGT features appeared in late 2025, in the VGGT-DP paper.

Licenses split inside both families

Neither family licenses as one thing. The split lines fall in different places.

In the DA3 repository table, read 28 October 2026, the small and base tiers carry Apache 2.0. The large, giant, and nested tiers carry CC BY-NC 4.0. That is not a commercial grant. The metric and monocular large models carry Apache 2.0. One row still conflicts between the publisher's own two sources. The repository table marks DA3-LARGE-1.1 non-commercial. That tier's own card says Apache 2.0. Both were read the same day. Until that reconciles, the safe commercial set is small, base, metric, and mono.

VGGT splits between code and weights. On 29 July 2025 the team moved the code to a license that allows commercial use. It excludes military use. The original VGGT-1B checkpoint stays CC BY-NC 4.0 on its card. Commercial rights attach to a separate checkpoint, VGGT-1B-Commercial. It was released 21 July 2025. The team says it performs the same. It sits behind an application gate. The application is processed automatically. It is still a gate. A product plan should treat that gate as a schedule risk, not a legal barrier.

The wider license question crosses model families. It has its own page, the GR00T license guide.

Where each model is served

Neither publisher sells calls to its own model. A thin hosting market has formed around both. All figures in this section are in US dollars. Each was published by the named seller on the date named.

For VGGT, fal hosts the 1B model as an image-to-JSON endpoint. It bills per compute second at 0.000975, read 28 October 2026. A community publisher serves a vggt-1b lane on Replicate. Same per-second rate, on an L40S machine. The page estimates a typical run near 0.0091, read the same day. There is also a free browser demo on the Hub. It runs with 475 likes. No first-party API exists.

For DA3, WaveSpeed sells image calls at a fixed 0.005 an image, re-read 28 October 2026. Community lanes for DA3 variants run on Replicate. One measured about 0.020 a run on the metric variant, re-read 28 October 2026. The vast.ai model-library row offers a ComfyUI template for the mono-large tier on rented boards. That is self-host packaging, not a served model. Its description says the wider DA3 family surpasses VGGT on multi-view depth and camera pose. That sentence is true for the giant tier. It is wrong for the mono tier it sits on. The mono tier does not do multi-view at all. Tier claims belong to tiers.

This platform serves the small tier under the da3-small model id. It has been on the public list since 6 October 2026. It is metered per image on the vision route. The full seller census for DA3, with the per-second lanes and the rental math, lives in the provider check page.

Pick by the job, not the name

The decision rules, keyed on real variables:

Single images and camera-rate loops. Run a DA3 small or base tier. Apache 2.0, the fastest rows in the speed table, and the best single-image scores of the three methods compared here. For a robot that needs depth every frame, 160.5 images per second against 34.1 is the whole argument. The pose gap does not enter a one-camera job.

Survey-grade multi-view pose, or dense rebuilding on one machine. VGGT still belongs on the shortlist. The DA3 small tier does not reach it. The tiers that beat it are non-commercial at giant size. The large tier is the honest middle. It is comparable to VGGT on pose and faster. Its 1.1 retrain is the row the repository now recommends.

Point tracks over time, or a frozen geometry backbone for a learned task. VGGT. Tracks are an output DA3 does not produce. The VGGT feature ecosystem, with robot policies built on it, is real and published.

Novel view synthesis from few views. DA3. Its Gaussian head predicts splats straight from the frozen backbone.

Shipping a commercial product. Read the license tables above first. Then read the hosting list. The Apache DA3 tiers and this platform's lane carry no application gate. VGGT commercial use runs through the gated checkpoint or the per-second endpoints. Read the seller's terms on the seller page, on the day you buy.

Both lines have moved since these two papers

The searcher's question names two 2025 papers. Both teams have shipped since.

The VGGT team released VGGT-Omega on 18 May 2026. Its project page announces it as the next step of VGGT. It simplifies the design to a single dense prediction head. Training draws on 15 times more supervised data. The reported gain is 77 percent on Sintel camera estimation over the previous best. Its checkpoint is application-gated. A CVPR 2026 Best Paper Finalist line sits on the page. A separate third-party paper, VGGT-World, extended VGGT toward autoregressive geometry in March 2026. That one is not from the VGGT team.

DA3's own line moved inside its repository. The refreshed 1.1 checkpoints were retrained after a training bug fix. The repository now recommends them over the original large and giant releases. Street scenes are called out as the big winner.

None of that retires the comparison. The checkpoints people actually download today are still VGGT-1B and the DA3 tiers. They stand at 515,230 and 234,136 all-time downloads on the Hub, read 28 October 2026. The successor papers raise the ceiling. They do not change the choice you make this month.

The hosted small tier here

The comparison above is between model families at source. This platform's part in it is one row. The small tier is served under the da3-small model id. It has been on the public list since 6 October 2026. It is metered per image. The lane gives you what the local small tier gives you. That means relative depth, confidence, and camera parameters, multi-view or single-view. It does not give metric depth. After the tables above, you know its pose accuracy sits below VGGT's. That is the honest row.

A first call:

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 with three file parts bills three images. A request refused before it is served bills nothing. The vision route page documents the whole contract. The run-vision reference carries every field. The current rate renders on the pricing page. It is drawn from the platform's model data:

An account opened on the console covers the lane. The small tier has its own model page with the full sheet. The DA3 small pricing guide works the budget math image by image. For pairing depth with masks on the same route, the robot vision segmentation and depth guide covers the two-model pattern.

Frequently asked questions

Is Depth Anything 3 better than VGGT?

It depends on the tier. In the DA3 paper's own tables, the giant tier beats VGGT on nearly every pose and reconstruction benchmark. The large tier is comparable. The small tier loses on all five pose sets. For single-image depth, DA3 beats VGGT clearly. A family-level yes or no is the wrong answer to a tier-level question.

Is VGGT free for commercial use?

The code is, since July 2025, under a license that excludes military use. The original checkpoint is not. It carries CC BY-NC 4.0. A separate commercial checkpoint exists. The team says it performs the same. You get it through an application gate. DA3 splits by tier instead. Small, base, metric, and mono are Apache 2.0 with no gate.

Is there a VGGT API?

Yes, two third-party ones, and no first-party one. fal hosts the 1B model per compute second. A community lane runs on Replicate per GPU second. Both were read 28 October 2026. The VGGT team publishes code, weights, and a free browser demo. It sells nothing.

Which one is faster?

On the DA3 paper's one-rig comparison, the small tier runs 160.5 images per second. VGGT runs 34.1. That is about 4.7 times faster. The small tier also holds roughly ten times more views in memory. The giant tier runs nearly level with VGGT, 37.6 against 34.1.

Which should I use for robotics?

Per-frame depth for a control loop: a DA3 small or base tier. You get speed, memory headroom, and Apache terms. Mapping or survey from many views: VGGT or the DA3 large tier. The small tier's pose accuracy sits below VGGT's on every benchmark measured. If your policy needs point tracks or frozen geometry features, VGGT is the published base for exactly that.

Is VGGT still the current model?

The team released VGGT-Omega in May 2026 as its successor. It is gated behind an application. The original VGGT-1B remains the most-downloaded, best-supported checkpoint of its line. A comparison made today, between the checkpoints people actually run, is still VGGT-1B against the DA3 tiers.

Last verified: 2026-10-28