ORAV-Bench/ORAV-Bench-Artifacts

ORAV Bench: Generation Outputs

1,852 generated videos · 48 recorded failures · 5 systems. Task prompts and canonical references are in ORAV-Bench.

benchmark.json
<model_id>/manifest.json
<model_id>/videos/<task_id>.mp4

benchmark.json binds the benchmark repository, fixed revision and input fingerprint once for this collection. Each model's manifest.json contains only config and the completed / failed task-ID lists. The folder name identifies the model; no model ID is repeated inside every record.

Completed videos use the fixed path <model_id>/videos/<task_id>.mp4; there is no repeated per-video path, hash or status object. Tasks absent from both lists were not submitted. A listed completed video must exist; a missing file is not treated as a generation failure.

To add a model, add its folder and manifest. For a separate collection, copy the matching benchmark.json. A different model version or generation configuration should use a distinct folder. Use scripts/build_manifest.py in the evaluation code to join selected manifests with the benchmark tasks.

Contributing model outputs

We welcome generation results from model providers and community contributors. Please use the original benchmark prompts and references, preserve task IDs, and include generation settings and failures in the manifest. To contribute, email the maintainers with the model name and a download link to outputs in the format above.

Generation settings

Kling settings were checked against the actual POST requests for all 380 selected outputs: sound=off in every request. For video inputs, keep_original_sound=yes is used for the 99 tasks with audio references and no for the other 242 video-input tasks; 39 image-only requests have no video-input setting. This input-audio preservation flag is separate from sound, and sound=off does not imply that every output file has no audio track. Requested durations come from the benchmark tasks.

Paper.