The dataset could not be loaded because the splits use different data file formats, which is not supported. Read more about the splits configuration. Click for more details.
Error code: FileFormatMismatchBetweenSplitsError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
WorldCrafter-Fast — Orbia Path A submission
Generated videos and self-evaluation results for all 720 Orbia cases (480 short, 240 long). This is a self-evaluated submission package, not an organizer-verified leaderboard entry.
Model: TencentARC/WorldCrafter. Benchmark: worldbench/Orbia.
Generation settings
- Official per-case scene captions, with no added first/third-person or camera-motion instructions.
- Input images resized directly to 640 x 384; no crop. Output is H.264 MP4, 16 FPS.
- 357 frames for short cases and 961 for long cases; padded generation tails removed.
- Camera-conditioned generation; requested relative camera trajectory and translation units retained.
- Fast native intrinsics: UCPE horizontal FOV 100 degrees, xi=0; RepEncoder nominal fx=fy=640, cx=320, cy=192. Dataset intrinsics are recorded in conditions but are NOT used to adapt Fast. Consequently this setup does not exactly reproduce the dataset camera intrinsics.
- 6 inference steps, CFG 1, seed 42, gravity alignment disabled.
- Fast weights: TencentARC/WorldCrafter-Fast revision d2ac12a654f8cafde4cb2e0ae7d5127ce2624939.
- Orbia dataset revision: 5edae28cca27eaed64a144882237ab70a539747e.
- Orbia evaluator commit: f7d8ddde3069a6d5ce75a87fa6b4fc896cc78a59.
- Evaluation settings and checkpoint choices follow that official evaluator; local paths are adapted. PyArrow 17.0.0 was pinned for compatibility with the evaluator's NumPy 1.x environment.
- Evaluation profile: group=world, control_profile=pose_and_action; all T1–T6 tiers evaluated.
Results
| Metric | Score |
|---|---|
| T1 | 85.788532 |
| T2 | 73.263235 |
| T3 | 78.205282 |
| T4 | 84.770282 |
| T5 | 70.717818 |
| T6 | 86.850484 |
| Overall | 80.502377 |
All scores were recomputed from the complete set of 720 per-case reports with the official aggregator. The rank fields in scores.json compare only models in this submission (one model), not the public leaderboard.
Files
videos/: one original generated MP4 per case; no re-encoding during packaging.conditions/: per-case camera/intrinsics policy, timeline and input hashes.results/scores/: globally aggregated scores.results/worldcrafter-fast/results/<case_id>/report.json: original evaluation reports.provenance/: original completion markers and pinned model/evaluator provenance.validation.json: hash, coverage, aggregation and official submission-check results.SHA256SUMS: checksums of the submission files, excluding this checksum file itself.
Historical runtime paths in the original reports/markers are provenance, not files required to open the videos. Original dataset inputs, reference images/depth, model weights, credentials and runtime logs are not included.
Validation
From the pinned Orbia checkout with its dataset available:
python tools/check_submission.py /path/to/this/repository --dataset /path/to/orbia-data
Attribution and use
This package is intended for non-commercial research and Orbia leaderboard verification. It derives from Orbia/Orbia, released under CC BY-NC-SA 4.0; source-specific terms and the benchmark's attribution requirements continue to apply. Model code and weights are available separately under their respective licenses.
- Downloads last month
- 549