PhyEduVideo

Last updated: Oct 7, 2026

On this page

Dimensions Data composition Metrics Leaderboard How to run

PhyEduVideo (IIIT Hyderabad, Adobe, IISER Thiruvananthapuram; WACV 2026) asks whether a T2V model can illustrate a teaching point, not whether it obeys a law taxonomy. The bundled Prompts/Prompts.json has 60 concepts in 6 categories (mechanics, oscillations and waves, fluids, thermodynamics, electromagnetism, optics), split into 205 visually descriptive teaching-point prompts.

That is education, not PhyGenBench (27-law commonsense generation), not PhyGround (13-law criteria on TI2V), and not PhyFPS-Bench-Gen (frame-rate audit).

Paper: PhyEduVideo. Project: meghamariamkm.github.io/phyeduvideo26. Code: meghamariamkm/PhyEduVideo.

WorldFoundry imports caller-supplied result files (JSON / JSONL / CSV), checks prompt/video coverage against the bundled manifests, and writes a scorecard. The InternVL / CLIP / VQA judge stack is not run in tree.

Official paper teaser: text-to-video failures on physics-education demonstrations.

Dimensions

Physics-education T2V
KindTask
Teaching-point illustration prompts (curriculum), not law-criteria grounding (PhyGround) and not PhyGenEval commonsense generation.
Mechanics
KindCategory
Bundled Prompts.json category (23 of 60 concepts).
Oscillations and waves
KindCategory
Bundled Prompts.json category.
Fluids
KindCategory
Bundled Prompts.json category.
Thermodynamics
KindCategory
Bundled Prompts.json category.
Electromagnetism
KindCategory
Bundled Prompts.json category.
Optics
KindCategory
Bundled Prompts.json category.

Data composition

Size
60 concepts in 6 bundled categories, decomposed into 205 teaching-point T2V prompts (Prompts/Prompts.json). WorldFoundry imports caller-supplied result files and can check prompt/video coverage. InternVL/CLIP/VQA judging is not run in tree.
Sources
github.com
Modalities
text · video
Notes
Result files are caller-supplied. Coverage checks pair prompts to videos; they do not validate how upstream scores were produced.

Metrics

Metric

semantic_adherence
DirectionHigher is better
Imported semantic match to the teaching-point prompt and rubric objects/actions.
physics_commonsense
DirectionHigher is better
Imported instructional physics-commonsense family (not PhyGenBench's PhyGenEval score).
motion_smoothness
DirectionHigher is better
Imported temporal motion-smoothness family.
temporal_flickering
DirectionHigher is better
Imported temporal-flickering family as exposed by the result file (higher is better on this surface).
phyeduvideo_average
DirectionHigher is better
Mean over the imported metric families.
PrimaryPhyEduVideo Average

Leaderboard

Results from the paper (not a live board; not WorldFoundry-run scores). Paper Average row, SA column. PhyEduVideo: A Benchmark for Evaluating Text-to-Video Models for Physics Education (2026).

Model	Average scientific accuracy (SA)
Wan2.1	0.83
PhyT2V	0.78
CogVideoX	0.75
Video-MSG	0.68
VideoCrafter2	0.62

How to run

worldfoundry-eval zoo benchmark-run \
  --benchmark-id phyeduvideo \
  --mode official-validation \
  --official-results-path <official_results.json> \
  --generated-artifact-dir <generated_videos> \
  --output-dir tmp/phyeduvideo/official-validation \
  --json

--run-official on the dedicated runner is the same import path for WorldFoundry-generated artifacts. Outputs include scorecard.json (primary phyeduvideo_average), results_summary.csv, and per-sample rows when the file exposes them.

← Benchmark Hub