World Labs Atlas Posts a 25.3 Average 3D Reconstruction Error, Ahead of Five Specialist Models: The Lead Comes From Indoors

World Labs released Atlas on September 1, 2026, describing it as an omni model pretrained from scratch to natively operate on text, images, video, and 3D. In the published 3D reconstruction table, Atlas records a mean absolute-relative pointmap error of 25.3, ahead of all five recent baselines, which span 28.7 to 47.7 on the same metric where lower is better. Unfolding the table dataset by dataset shows the lead concentrated on one side. On the indoor NRGBD dataset Atlas posts 6.5 against 13.3 for the strongest baseline, more than a twofold gap, while on the driving dataset KITTI the two sit at 60.0 and 60.2 and on Tanks and Temples both land on exactly 42.4. ASAP works only from the table and figures in World Labs' official post to separate where Atlas actually wins from where it does not yet.

Behind the 25.3 average sits a split report card between indoors and outdoors

Atlas reconstruction results tell two different stories depending on whether they are read as one average or as seven datasets. World Labs reports mean absolute-relative pointmap error per dataset, where lower is better. Atlas averages 25.3 against Pi3X at 28.7, π³ at 34.7, VGGT-Ω 1B at 36.4, Depth Anything 3 at 39.3, and MapAnything at 47.7.

The gap is distributed unevenly across datasets. On NRGBD, an indoor capture set, Atlas records 6.5 while the strongest baseline Pi3X records 13.3, a gap of more than two to one. ETH3D shows 9.3 against 18.7, again roughly double; the object-scan dataset DTU shows 8.6 against 11.1, and the indoor scan set ScanNet shows 12.4 against 15.7. Two entries run the other way. On the driving dataset KITTI, Atlas posts 60.0 and Pi3X posts 60.2, a difference of 0.2, and on Tanks and Temples both models land on 42.4 to the first decimal.

That distribution is a clue about what this model learned. The datasets where Atlas wins by a wide margin involve a camera circling inside a bounded space and revisiting the same surfaces from several angles. The datasets where it merely ties involve a camera traveling in one direction through driving scenes and large outdoor structures, never returning to the same point. A design that takes 3D as a native input and carries spatial context pays off where a scene can be looked at again, and hits the same wall as specialist models where a pass is one-way. Indoor absolute errors between 6.5 and 12.4 against 60.0 on KITTI shows how high that wall still stands.

One model producing seven kinds of output is the core of this release

The structural claim in Atlas is not a per-task ranking but the consolidation of video, 3D, and panorama into seven output types from a single model. World Labs describes Atlas as a multimodal autoregressive diffusion transformer, pretrained from scratch on text, images, video, and 3D simultaneously rather than a video model with 3D bolted on. It treats precise camera geometry as a native input type and grounds inputs in 3D space to form spatial context.

The input and output range backs that claim. Camera-controlled generation takes one to six input images, spatial reconstruction takes one to dozens, and faithful recreation can consume over 100 input images. Outputs include up to one minute of video at 1440p, point clouds, 3D Gaussian splats, 360-degree panoramas from text or image prompts, and RGB plus depth data for robotic sensor simulation. For robotics applications, 24 frames capture an environment.

What this means in practice is pipeline collapse. Turning one scene into video, then into 3D geometry, then into a robot simulation has required stitching a generation model, a reconstruction model, and a simulator together while reconciling coordinate frames and scale between them. Error accumulated at each conversion, and changing an upstream stage forced a re-fit downstream. A single model emitting seven output types over one shared spatial representation removes those seams. That every comparison in the reconstruction table is a reconstruction specialist points the same way: a general model clearing the specialists' average is offered as evidence that consolidation was not paid for in accuracy.

Preference ranges from 75% to 94%, a 19-point spread depending on the opponent

Atlas was preferred over all five comparison models on camera-controlled generation by third-party human evaluators. The rates at which Atlas was chosen are 75% against MiniMax H, 81% against Gemini Omni Flash, 86% against Happy Horse 1.1, 93% against FLUX 3, and 94% against Seedance 2.5. The spread across those five values is 19 percentage points.

The number that matters in a win-rate table is the floor, not the ceiling. A 94% rate is achievable against an opponent that never targeted camera control in the first place, while the figure that reflects real competitive pressure is 75%. Against MiniMax H, Atlas was chosen three times out of four, which also means evaluators picked the other side once out of four. Capable alternatives already exist for moving a camera exactly as specified.

The two evidence types also need to be kept apart. The error table is an absolute metric measured against ground-truth 3D on datasets, while preference is a relative metric from humans comparing two outputs side by side. The latter measures which output looks better rather than which is more accurate, and the two do not always produce the same ordering. World Labs using error tables for reconstruction and preference for generation follows the convention for each task, but the two numbers should not be read in one breath.

What is disclosed and what is not

The boundary between what is verifiable about Atlas today and what is not is sharp. Verifiable: the architecture family, the input and output specifications, two kinds of evaluation results, and the distribution status. Atlas is entering early access with select partners through a request form, and will power future versions of Marble, World Labs' creation tool.

A substantial list is missing. There is no pricing information and no general availability date. Model size, training data scale, the time and compute required to produce one minute of 1440p video are all undisclosed. Whether real-time interaction is possible, or only batch generation, is not stated. Nor is there any evaluation of whether the RGB and depth output for robot simulation is physically consistent enough to train policies on.

The comparison conditions in the reconstruction table have gaps too. One of the five baselines is labeled Pi3X (posed), indicating it was measured with camera poses supplied, but the table alone does not settle whether Atlas was measured under the same condition. Confirming the exact protocol requires a document at technical-report depth.

What a team can take from this announcement today

For robotics and spatial computing teams, the immediately usable part of the Atlas announcement is the evaluation design rather than model access. With early access limited to select partners, most teams cannot run this model now, but the way World Labs chose to measure it transfers directly: reconstruction scored across indoor, outdoor, and object-scan datasets separately instead of a single average, with generation quality measured separately through human preference.

The contrast between KITTI at 60.0 and NRGBD at 6.5 in particular gives a way to place your own task. For work that captures an indoor space from multiple angles and rebuilds it in 3D, the unified-model line has already reached a usable range. For work that captures in one pass from a vehicle or a drone, no model including Atlas has meaningfully reduced the error yet, so redesigning the capture path to allow revisiting beats swapping models.

The robotics specification of 24 frames per environment is also usable as a planning baseline. If the minimum capture needed to hand an environment to the model is on the order of 24 frames, teams that have been budgeting hundreds of images per scene have room to recalculate. Running a collection pilot against that baseline now, independent of the early access request, means the day access opens the verification can start on your own data.

Source: ASAP analysis based on World Labs' official Atlas post (September 1, 2026), including its reconstruction benchmark table, human preference evaluation, and model specifications