WorldFoundry
English中文

PhyGround

Edit this pageGitHubMarkdown

On this page
DimensionsData compositionMetricsLeaderboardHow to run

PhyGround (Northeastern University) is criteria grounding, not filmed continuation and not PhyGenEval. Each of 250 text+image-to-video prompts names an expected physical outcome and 2–4 applicable laws from a 13-law taxonomy (solid-body, fluids, optics). A judge (PhyJudge-9B or humans) rates the generated video 1–5 on semantic adherence, physical temporal validity, persistence, and each applicable law. That is closer to a rubric than to Physics-IQ mask IoU.

PhyGenBench is the T2V generation commonsense suite (160 prompts / 27 laws / PhyGenEval). PhyGround is TI2V law-criteria scoring on 250 prompts / 13 laws.

Paper: PhyGround. Project: phyground.github.io. Code: NU-World-Model-Embodied-AI/PhyGround. Data: NU-World-Model-Embodied-AI/phyground.

WorldFoundry imports scores.json from PhyJudge/vLLM or a cloud judge. It does not execute PhyJudge-9B over candidate videos.

Official paper teaser: physical-law violations in generative world models.
Official paper teaser: physical-law violations in generative world models.

Dimensions

  • Criteria-grounded TI2V
    KindTask

    Text+image-to-video prompts with an expected physical outcome and 2–4 applicable laws. Videos are judged 1–5 on criteria — not pixel-matched to a filmed future.

  • Solid-body mechanics
    KindCategory

    Domain score over applicable solid-body laws in the 13-law taxonomy.

  • Fluid dynamics
    KindCategory

    Domain score over applicable fluid-dynamics laws.

  • Optics
    KindCategory

    Domain score over applicable optics laws.

Data composition

  • Size

    250 text+image-to-video prompts over a 13-law taxonomy (solid-body, fluids, optics). WorldFoundry imports PhyJudge/human scores.json; it does not run PhyJudge-9B.

  • Sources

    NU-World-Model-Embodied-AI/phyground · github.com

  • Modalities

    image · text · video

  • Notes

    Judge outputs live under annotations/ or --official-results-path. Use --result-model-id when a file holds several model ids.

Metrics

Metric

  • semantic_adherence
    DirectionHigher is better

    1–5 semantic alignment with the prompt, then normalized.

  • physical_temporal_validity
    DirectionHigher is better

    1–5 physical temporal validity of the generated event.

  • persistence
    DirectionHigher is better

    1–5 object/state persistence through the clip.

  • solid_body_score
    DirectionHigher is better

    Aggregate over applicable solid-body laws for that video.

  • fluid_dynamics_score
    DirectionHigher is better

    Aggregate over applicable fluid-dynamics laws.

  • optics_score
    DirectionHigher is better

    Aggregate over applicable optics laws.

  • phyground_overall
    DirectionHigher is better

    Imported overall after general dimensions and domain scores.

    PrimaryPhyGround Overall

Leaderboard

Results from the paper (not a live board; not WorldFoundry-run scores). PhyGround: Benchmarking Physical Reasoning in Generative World Models (2026).

ModelOverall (1–5 human)
Wan2.2-27B-A14B3.28
Veo-3.1closed3.28
OmniWeaving3.10
Cosmos-14B2.91
LTX-2.3-22B2.69
Wan2.2-TI2V-5B2.63
Cosmos-2B2.58
LTX-2-19B2.56

How to run

Stage the data root (prompts/, first_images/, annotations/, videos/<model_id>/) and import judge outputs:

hf download NU-World-Model-Embodied-AI/phyground --repo-type dataset --local-dir <phyground_data_root>

worldfoundry-eval zoo benchmark-run \
  --benchmark-id phyground \
  --mode official-validation \
  --benchmark-data-root <phyground_data_root> \
  --generated-artifact-dir <phyground_data_root>/videos/<model_id> \
  --result-model-id <model_id> \
  --output-dir tmp/phyground/official-validation \
  --json

Or pass --official-results-path at the scores.json / annotations/ file. --prompt-manifest optionally checks prompt coverage. Outputs include scorecard.json (primary phyground_overall).

← Benchmark Hub

Related

OverviewBenchmark HubMetricsScorers & qualityEnvironmentCLILocal assets