Visual Chronometer

Visual Chronometer (TACO Group, Texas A&M) predicts Physical Frames Per Second (PhyFPS) — the temporal scale implied by visual motion — ignoring container metadata. It is trained with controlled temporal resampling and physics-grounded augmentations (motion blur, rolling shutter). As a WorldFoundry benchmark it scores a directory of arbitrary videos (not a fixed prompt suite) with mean PhyFPS plus intra- / inter-video coefficients of variation. Lower values mean steadier temporal behavior.

Sibling generation audit: PhyFPS-Bench-Gen (100-prompt Meta-vs-PhyFPS).

Code: taco-group/Visual_Chronometer. Checkpoint: xiangbog/Visual_Chronometer (vc_common_10_60fps.ckpt, 10–60 FPS).

WorldFoundry runs predict.py in-process through the shared PhyFPS-Bench-Gen wrapper, or imports results.csv. Runtime / normalizer surfaces exist; no release-verified full official run is claimed.

Dimensions

• PhyFPS prediction

  KindAxis

  Sliding-window physical FPS implied by visual motion. All consistency metrics lower-is-better.

Data composition

• Size

  TACO PhyFPS predictor over arbitrary videos (not a fixed prompt suite). Released VC-Common checkpoint covers 10–60 FPS. Shared runtime with PhyFPS-Bench-Gen.

• Sources

  github.com

• Modalities

  video

• Notes

  Runtime/normalizer surfaces exist; no release-verified full official run is claimed. Mock backends are wiring-only.

Metrics

Metric

• mean_phyfps

  DirectionLower is better

  Mean average PhyFPS across evaluated videos.

• inter_video_cv

  DirectionLower is better

  Cross-video coefficient of variation over average PhyFPS.

• intra_video_cv

  DirectionLower is better

  Mean within-video coefficient of variation over sliding-window clips.

• visual_chronometer_average

  DirectionLower is better

  Mean over available Visual Chronometer consistency score families.

  PrimaryVisual Chronometer Average

Leaderboard

Results from the paper (not a live board; not WorldFoundry-run scores). Table 2 predictor accuracy, not a generator ranking. Video-based VLM rows. The Pulse of Motion: Measuring Physical Frame Rate from Visual Dynamics (2026).

Model	MAE on PhyFPS-Bench-Real (lower is better)
VC-Commonours	3.46
VC-Wideours	7.76
Seed-1.6-Flashvideo VLM	20.00
Seed-1.6video VLM	20.40
Gemini-3.1-Provideo VLM	21.67
Gemini-3-Flashvideo VLM	23.40

How to run

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

--generated-artifact-dir plus WORLDFOUNDRY_VISUAL_CHRONOMETER_PREDICT_BACKEND selects local prediction. Env: WORLDFOUNDRY_VISUAL_CHRONOMETER_RESULTS_PATH, WORLDFOUNDRY_VISUAL_CHRONOMETER_ROOT.

In-tree pointers: worldfoundry/data/benchmarks/tasks/external/visual-chronometer.yaml, worldfoundry/evaluation/tasks/execution/runners/phyfps_bench_gen/runtime/visual_chronometer.

Outputs include scorecard.json (primary visual_chronometer_average) and results.csv.

← Benchmark Hub