CornerBench-1200

A collection of 1,200 generated object / scene editing cases with synchronized front-left, front, and front-right camera videos, for 3,600 camera videos in total.

This is a generated-video collection built on nuPlan driving scenes. A case is one generated example, not necessarily a unique underlying recorded scene. The cases are not assigned to independent train/validation/test splits.

Contents

Property	Value
Resolution per camera	1280 × 720
Frame rate	10 fps
Video encoding	H.264 in MP4, no audio
Object / scene editing clip lengths	25–69 frames (2.5–6.9 s)
Video payload	14,264,240,964 bytes
Cameras per case	left, front, right

videos/
  object_0000/{left,front,right}.mp4
  ...
  object_1199/{left,front,right}.mp4
thumbnails/             # one front-camera preview per case
metadata.jsonl          # one record per case; relative video paths
manifest.json           # detailed video dimensions, timing, sizes and hashes
SHA256SUMS              # SHA-256 checksums for all 3,600 videos
gallery/                # portable static three-view explorer

Download and inspect

from huggingface_hub import snapshot_download
import json
from pathlib import Path

root = Path(snapshot_download(
    repo_id="cornerbench-anon-2027/CornerBench-1200",
    repo_type="dataset",
))
manifest = json.loads((root / "manifest.json").read_text())
example = manifest["cases"][0]
front_video = root / example["videos"]["front"]["path"]
print(example["case_id"], front_video)

Each metadata.jsonl row contains a case ID, case family, category, motion, direction, frame count, frame rate, clip duration, thumbnail path, and three relative video paths. The detailed manifest includes a SHA-256 digest for every video. left, front, and right are path strings, not embedded video bytes.

The static explorer synchronizes the three videos with shared play/pause, frame stepping, timeline seeking, and playback speed, and supports category search and pagination. The gallery's public version fetches its media from this dataset repository.

Validation and scope

All 3,600 videos were fully decoded without errors. Camera dimensions, frame rates, and frame counts agree within each case. Release packaging removes inherited container metadata without re-encoding the H.264 video stream. Five object-editing cases use their corrected side-camera renderings.

This release contains generated observations and basic descriptive metadata. It does not include a complete closed-loop simulator, planner trajectories, source logs, calibration files, or an independent safety validation. Frame counts and camera dimensions are checked; visual correctness and physical consistency are not guaranteed by those checks.

Provenance, attribution, and terms

The source driving data comes from nuPlan, provided by Motional. The videos have been modified through generated object/scene edits, and are not original unmodified nuPlan footage.

nuPlan is distributed under CC BY-NC-SA 4.0 with additional nuPlan terms, including noncommercial restrictions. The full upstream terms are retained in LICENSE-nuPlan.txt, obtained from the official nuPlan license. Consult those terms together with the upstream data attribution; this release does not remove or supersede them. The metadata uses a custom license label so that it does not imply unqualified CC BY-NC-SA licensing without the additional terms.

The generated videos use the CornerFactory/NeoVerse pipeline. Related controller weights are grouped in the separate CornerFactory model repository. No NVIDIA PhysicalAI training data is included in this dataset.