Submit to the Leaderboard

Submit your model to the WorldAtlas leaderboard by emailing jiadonghou3@gmail.com with a download link to your submission.

Submission options

Option	What to submit
Self-evaluation	Generated videos, generation records, configs, and the complete WorldAtlas evaluation report.
Evaluation by our team	Generated videos, generation records, and configs. The WorldAtlas team runs the evaluation; no local metric setup or scores are required.

For setup and generation, follow Getting Started and Run Evaluation. Self-evaluation uses the same published benchmark configuration and model-specific overrides as the corresponding leaderboard track.

Tracks and coverage

Track	Required outputs
image_static	Videos for the static-image tasks.
image_dynamic	Videos for the dynamic-image tasks.
memory_loop	Closed-loop rollouts, submitted separately for 60s and 30s.

Choose the tracks supported by your model and cover every selected case in the released manifest. Include the input manifest used for generation. A run with --limit is a smoke run, not a complete leaderboard submission. video_static and video_dynamic provide reference videos for scoring; they are not generation tracks.

Preserve the prescribed input images, prompts, masks, and camera/action controls. Keep the original generated videos and their frame rate, resolution, and duration; disclose any post-processing.

Prepare the submission

Use one folder per track and memory duration:

my-model/
  meta.json
  configs/
  environment.txt
  image_static/
    input_manifest.jsonl
    predictions/
    reports/

• predictions/: copy the complete generation output directory, including videos, evaluation_manifest.jsonl, generation_records.jsonl, generation_manifest.json, prompt_map.json, and any camera/action sidecars. Preserve relative paths. For sharded runs, include every shard’s manifests and records.
• reports/: for self-evaluation, include the complete evaluation output directory: run_manifest.json, per_sample_scores.jsonl, per_metric_summary.json, per_dimension_summary.json, leaderboard.csv, and any additional distribution, failure, or applicability reports. For team evaluation, omit this folder.
• configs/ and environment.txt: include the model and benchmark configs actually used, inference settings and seed, code/weight versions, hardware, and package versions. Include the evaluation environment for self-evaluation.

For videos generated outside WorldAtlas, provide a mapping from each released sample_id to its video, along with the source inputs, prompts/controls, and generation records. The team will confirm the input-to-video binding before evaluation. Renaming videos alone does not create a bound evaluation manifest.

Add a meta.json describing the submission; this is metadata for the team to review:

{
  "model_name": "my-model",
  "display_name": "My Model",
  "organization": "Your Lab",
  "contact_email": "you@example.org",
  "submission_mode": "team_evaluation",
  "tracks": [{"suite": "image_static"}],
  "worldatlas_commit": "<Git commit used for generation>",
  "dataset_revision": "<Hugging Face dataset revision>",
  "model_url": "https://example.org/my-model",
  "notes": ""
}

Use self_evaluation for submissions with reports. A memory-track entry is {"suite": "memory_loop", "duration_seconds": 60} (or 30). Add paper, code, and weights links when available, and describe any deviations from the published protocol in notes.

Send and review

Upload the folder to Hugging Face, Google Drive, or another host accessible to the team. Email [WorldAtlas Submission] <model-name> with the download link and attach meta.json. For restricted links, include access instructions.

The team checks coverage, input/output correspondence, and evaluation settings; runs evaluation when requested; and follows up on missing materials or failed cases before adding reviewed results to the leaderboard. List any known failed or missing cases in the submission.