LikePhys

Last updated: Oct 7, 2026

On this page

Dimensions Data composition Metrics Leaderboard How to run

LikePhys (University of Oxford, ICLR 2026) is a training-free likelihood probe, not a generation bench. It does not score videos you generate. For each of 12 simulated scenarios (rigid bodies, fluids, deformables, optics), it pairs a physically valid clip with matched implausible variations and asks whether the diffusion model assigns higher likelihood (lower denoising ELBO) to the valid clip. The primary metric is mis-rank rate — how often the model prefers the implausible variation. Lower is better everywhere.

That is the opposite job from Physics-IQ (continue a filmed scene) and from PhyGenBench (judge a generated T2V clip).

Paper: LikePhys. Project: yuanjianhao508.github.io/LikePhys. Code: YuanJianhao508/LikePhys. Clips: JianhaoDYDY/LikePhys-Benchmark (920 clips).

WorldFoundry recomputes the official mis-rank formulas in tree (compute_misrank_normalized + read_exp_final.py filters). Driving the upstream ELBO probe needs a caller-supplied LikePhys checkout (WORLDFOUNDRY_LIKEPHYS_EVALUATOR_ROOT), CUDA, and the probed checkpoints — WorldFoundry does not vendor the forked diffusers pipelines.

Official paper teaser: a model with learned physics should assign higher likelihood to valid videos than to violations.

Dimensions

Model likelihood probe
KindTask
Scores the diffusion model itself (denoising ELBO on paired clips). There is no candidate-video input and no generation leaderboard.
Rigid body
KindCategory
Ball drop, ball collision, pendulum, block slide, and pyramid.
Fluids
KindCategory
Droplet, faucet, and river scenarios.
Deformable
KindCategory
Cloth drape and waving flag.
Optics
KindCategory
Moving-light and moving-camera shadow scenarios.

Data composition

Size
12 simulated scenarios × 10 subgroups; catalog records 920 clips (physically valid clip + matched implausible variations). Clips come from JianhaoDYDY/LikePhys-Benchmark. WorldFoundry recomputes official mis-rank; the ELBO probe needs a caller-supplied LikePhys checkout.
Sources
JianhaoDYDY/LikePhys-Benchmark · github.com
Modalities
text · video
License
apache-2.0
Notes
Mis-rank is an error rate (lower is better). Paper filter_config exclusions apply unless --no-report-filter is set.

Metrics

Metric

likephys_misrank_rate
DirectionLower is better
Variation-weighted mean mis-rank rate over every scored scenario/variation pair (published ranking).
PrimaryLikePhys Mis-Rank Rate
likephys_dataset_weighted_misrank_rate
DirectionLower is better
Mean over per-scenario mis-rank rates, weighting each of the 12 scenarios equally.
rigid_body_misrank_rate
DirectionLower is better
Mis-rank over ball drop, ball collision, pendulum, block slide, and pyramid.
fluid_misrank_rate
DirectionLower is better
Mis-rank over droplet, faucet, and river.
deformable_misrank_rate
DirectionLower is better
Mis-rank over cloth drape and waving flag.
optics_misrank_rate
DirectionLower is better
Mis-rank over moving-light and moving-camera shadow scenarios.
temporal_disorder_misrank_rate
DirectionLower is better
Mis-rank restricted to the temporal_disorder variation shared by all 12 scenarios.

Leaderboard

Results from the paper (not a live board; not WorldFoundry-run scores). Paper overall PPE. Lower means stronger intuitive physics. LikePhys: Evaluating Intuitive Physics Understanding in Video Diffusion Models via Likelihood Preference (2025).

Model	PPE average (lower is better)
Hunyuan T2V	43.6
Wan2.1-T2V-14B	43.8
CogVideoX1.5–5B	43.8
LTX v0.9.5	44.7
Wan2.1-T2V-1.3B	48.0
CogVideoX-2B	48.2
CogVideoX-5B	49.8
Mochi	51.9

How to run

Import results_<model>.json files (or a directory of <scenario>/results_<probe_model>.json) and recompute mis-rank:

worldfoundry-eval zoo benchmark-run \
  --benchmark-id likephys \
  --mode official-validation \
  --official-results-path <likephys_results_root> \
  --output-dir tmp/likephys/official-validation \
  --json

Use --probe-model when a results root holds several backends. --no-report-filter keeps variations the paper's filter_config excludes.

Run the upstream ELBO probe for one model on one scenario (CUDA + LikePhys checkout):

hf download JianhaoDYDY/LikePhys-Benchmark --repo-type dataset --local-dir "${WORLDFOUNDRY_LIKEPHYS_DATA_ROOT}"

WORLDFOUNDRY_LIKEPHYS_EVALUATOR_ROOT=/path/to/LikePhys \
${WORLDFOUNDRY_UNIFIED_PYTHON} \
  worldfoundry/evaluation/tasks/execution/runners/likephys/run_likephys_official_runner.py \
  --run-official \
  --probe-model wan2.1-T2V-1.3b \
  --scenario ball_drop \
  --benchmark-data-root "${WORLDFOUNDRY_LIKEPHYS_DATA_ROOT}" \
  --output-dir tmp/likephys/probe \
  --json

--dry-run prepares the workspace and prints commands. Outputs include scorecard.json (primary likephys_misrank_rate) and misrank_breakdown.json.

← Benchmark Hub