patina27/atlas-bench

Atlas Bench

A unified, camera-accurate benchmark built from every result World Labs published in its Atlas blog post: the input images, the camera of every requested view or video frame, and Atlas's own output for each. Use it to run a novel-view-synthesis / camera-controlled video model on exactly the same inputs and cameras as Atlas and compare.

Browse everything in the viewer Space: patina27/atlas-bench-viewer (3D camera view, input frusta, video playback synced with the camera path).

Contents

track	tasks	what
camera_control	20	1 input image -> 96 views on a sphere around the scene
diverse_paths	3	8-24 posed photos of a real scene -> 3-4 long camera-path videos each (11 paths)
long_video	4	a few posed images -> 1-minute videos along designed camera paths
multi_image	2	3 / 25 real photos (poses estimated by us) -> Atlas's 1/2/3- and 2/6/25-photo reconstructions
showcase_*	198	browse-only sections without usable cameras (spatial-context stitching, reframing, robotics, point clouds, Gaussian splats, panoramas, text-to-image)

See FORMAT.md for the file layout. Every task has scene.json (cameras exactly as published) and, where cameras exist, cameras_opencv.json (OpenCV camera-to-world + pixel K, plus K for an 832x480 center crop).

Cameras

The blog's camera convention is not documented; we verified it on the published frames (epipolar error of Atlas's own frames under the published cameras, ~0.2-0.4 px; every alternative convention gives 8-40 px). Details, per-track numbers and figures: docs/CONVENTION.md. In short: position = camera center, quaternion = [x, y, z, w] camera-to-world rotation of a three.js / OpenGL camera; OpenCV c2w = [R(q) diag(1, -1, -1) | position]. tools/atlas_bench/cameras.py implements the conversions (tested).

Known caveats:

• diverse_paths and long_video inputs are ultra-wide (living room ~129 deg horizontal FOV) and keep some lens distortion; residuals ~2 px.
• showcase_highlight/robot_orbit_v4 and world_labs_atlas_v3 input poses are unreliable.
• multi_image photo poses are our estimates, registered to Atlas's world (all 3 house-garden photos; only 2 of 25 Stanford photos). Method and accuracy: docs/MULTI_IMAGE_POSES.md.

Loading

import json, numpy as np
from huggingface_hub import snapshot_download

root = snapshot_download("patina27/atlas-bench", repo_type="dataset")
scene = json.load(open(f"{root}/tasks/diverse_paths/studio-loft/scene.json"))
cams = json.load(open(f"{root}/tasks/diverse_paths/studio-loft/cameras_opencv.json"))
c2w = np.array(cams["targets"]["path-1"]["c2w"])     # (400, 4, 4) OpenCV camera-to-world
K = np.array(cams["targets"]["path-1"]["K"])         # (400, 3, 3) at the video's file resolution

SHA256SUMS covers every file; VERSION is the dataset version.

Source

All media and cameras come from the public World Labs blog and its CDN (wlt-ai-cdn.art); manifests/ keeps the page manifests they were extracted from and every scene.json records its source URLs. The OpenCV cameras, the multi_image photo poses and the documentation are ours.