physics-from-video/morpheus-real-world

Morpheus — Real-World Physics Videos

Real-world reference footage for Morpheus, a benchmark that tests whether video generative models (Wan, CogVideo, LTX-Video, COSMOS-predict1/2, Pyramid-Flow, Veo3, Kling-Turbo, ...) obey Newtonian mechanics. Object trajectories are extracted via SAM2 tracking and tested against physical laws (energy/momentum conservation, equations of motion) rather than pixel-matched to a single "correct" video.

This repo contains the filmed real-world physics experiments used as the ground-truth reference split in the paper. Code to download, track, and score this data lives in physics-from-video/Morpheus.

Quickstart

pip install morpheus-physics   # or: pip install -e . from the Morpheus repo
export HF_TOKEN=hf_...         # optional — this dataset is public, only helps with rate limits
morpheus-download-data --local-dir ./data
morpheus-track-and-score --input-dir ./data --output-dir ./results \
    --methods real-world --calculate-scores

Directory layout

real-world-cropped/<experiment>/<video_dir>/
    cropped_video.mp4          # the filmed clip, resized to 1280x1024
    info.json                  # experiment / model / path metadata
    frames_for_tracking/       # per-frame JPGs used for SAM2 tracking

Experiments covered (16)

These are exactly the phenomenon buckets reported in the paper's real-world results tables (REAL_WORLD_DEFAULT_EXPERIMENTS in morpheus/taxonomy.py):

Phenomenon bucket	Experiments
Free fall	falling_ball, falling_apple, falling_marker, falling_tape
Projectile motion	projectile
Bouncing	bouncing_ball
Sliding / rolling object	sliding_book, rolling_full_can, rolling_empty_can, rolling_orange
Pendulum	holonomic_pendulum
Double pendulum	double_pendulum
Spring-mass	spring
Collision	collision_equal, collision_big_hits_small, collision_small_hits_big

124 videos total. (The pipeline can also score a non_holonomic_pendulum experiment, but it is intentionally excluded here — the paper's results tables omit it, so it is not part of this release.)

Citation

@inproceedings{tragoudaras2026evaluating,
  title={Evaluating Newtonian Mechanics in Video Generative Models with Real Physical Systems},
  author={Antonios Tragoudaras and Chenyu Zhang and Daniil Cherniavskii and Antonios Vozikis and Thijmen Nijdam and Derck W. E. Prinzhorn and Mark Bodracska and Nicu Sebe and Andrii Zadaianchuk and Stratis Gavves},
  booktitle={Forty-third International Conference on Machine Learning},
  year={2026},
  url={https://openreview.net/forum?id=f4xAzcMug6}
}