P
Physics World Model

Demos

Canonical demo datasets — the same scenes used in the InverseNet paper: 10 KAIST-10 hyperspectral scenes for CASSI and the 6 standard SCI Video Benchmark videos (Kobe, Traffic, Runner, Drop, Crash, Aerial) for CACTI. Each sample is 256×256, deterministic, and committed to the repo so external users can verify reconstruction byte-for-byte. Download the input, run the reference solver locally, and see the PWM pipeline work end-to-end — or compete with a better solver.

T1_nominal

cacti

L3-004 (CACTI) — InverseNet SCI Video Benchmark

Full benchmark page → L3-004
reference PSNR
28.0 dB
your solver should beat this
Shape summary
snapshot: [256, 256]
ground truth: [8, 256, 256]
reference solution: [8, 256, 256]
sample_01 · kobe
ref PSNR 27.99 dB
Measurement
sample_01 Measurement
Ground truth
sample_01 Ground truth
Reference
sample_01 Reference
snapshot.npz354.6 KBmeta.json1.3 KBsolution.npz1.79 MBground_truth.npz1.72 MBground_truth.png28.6 KBsolution.png30.6 KBsnapshot.png39.4 KB
sample_02 · traffic
ref PSNR 20.88 dB
Measurement
sample_02 Measurement
Ground truth
sample_02 Ground truth
Reference
sample_02 Reference
snapshot.npz256.2 KBmeta.json1.3 KBsolution.npz1.75 MBground_truth.npz548.3 KBground_truth.png38.1 KBsolution.png38.3 KBsnapshot.png47.5 KB
sample_03 · runner
ref PSNR 29.63 dB
Measurement
sample_03 Measurement
Ground truth
sample_03 Ground truth
Reference
sample_03 Reference
snapshot.npz250.7 KBmeta.json1.3 KBsolution.npz1.73 MBground_truth.npz429.7 KBground_truth.png30.7 KBsolution.png31.6 KBsnapshot.png38.0 KB
sample_04 · drop
ref PSNR 34.83 dB
Measurement
sample_04 Measurement
Ground truth
sample_04 Ground truth
Reference
sample_04 Reference
snapshot.npz254.4 KBmeta.json1.3 KBsolution.npz1.66 MBground_truth.npz254.7 KBground_truth.png17.1 KBsolution.png23.4 KBsnapshot.png36.9 KB
sample_05 · crash
ref PSNR 25.73 dB
Measurement
sample_05 Measurement
Ground truth
sample_05 Ground truth
Reference
sample_05 Reference
snapshot.npz311.2 KBmeta.json1.3 KBsolution.npz1.76 MBground_truth.npz816.6 KBground_truth.png29.0 KBsolution.png31.4 KBsnapshot.png41.6 KB
sample_06 · aerial
ref PSNR 25.06 dB
Measurement
sample_06 Measurement
Ground truth
sample_06 Ground truth
Reference
sample_06 Reference
snapshot.npz332.7 KBmeta.json1.3 KBsolution.npz1.75 MBground_truth.npz917.8 KBground_truth.png31.4 KBsolution.png31.3 KBsnapshot.png42.4 KB
Run it
python3 scripts/regenerate_demos_inversenet.py --only cacti (uses pwm_core.recon.cacti_solvers.gap_tv_cacti)
T1_nominal

cassi

L3-003 (CASSI) — InverseNet KAIST-10 benchmark

Full benchmark page → L3-003
reference PSNR
35.3 dB
your solver should beat this
Shape summary
snapshot: [256, 310]
ground truth: [256, 256, 28]
reference solution: [256, 256, 28]
sample_01 · scene01
ref PSNR 35.3 dB
Measurement
sample_01 Measurement
Ground truth
sample_01 Ground truth
Reference
sample_01 Reference
snapshot.npz391.4 KBmeta.json1.3 KBsolution.npz6.22 MBground_truth.npz6.19 MBground_truth.png23.8 KBsolution.png20.0 KBsnapshot.png32.0 KB
sample_02 · scene02
ref PSNR 36.13 dB
Measurement
sample_02 Measurement
Ground truth
sample_02 Ground truth
Reference
sample_02 Reference
snapshot.npz390.9 KBmeta.json1.3 KBsolution.npz6.31 MBground_truth.npz6.30 MBground_truth.png25.9 KBsolution.png25.8 KBsnapshot.png33.6 KB
sample_03 · scene03
ref PSNR 35.66 dB
Measurement
sample_03 Measurement
Ground truth
sample_03 Ground truth
Reference
sample_03 Reference
snapshot.npz391.0 KBmeta.json1.3 KBsolution.npz6.28 MBground_truth.npz6.24 MBground_truth.png18.4 KBsolution.png19.7 KBsnapshot.png36.5 KB
sample_04 · scene04
ref PSNR 40.05 dB
Measurement
sample_04 Measurement
Ground truth
sample_04 Ground truth
Reference
sample_04 Reference
snapshot.npz388.3 KBmeta.json1.3 KBsolution.npz6.16 MBground_truth.npz6.22 MBground_truth.png15.4 KBsolution.png13.8 KBsnapshot.png23.3 KB
sample_05 · scene05
ref PSNR 32.84 dB
Measurement
sample_05 Measurement
Ground truth
sample_05 Ground truth
Reference
sample_05 Reference
snapshot.npz389.9 KBmeta.json1.3 KBsolution.npz6.22 MBground_truth.npz6.21 MBground_truth.png19.8 KBsolution.png19.7 KBsnapshot.png32.0 KB
sample_06 · scene06
ref PSNR 34.56 dB
Measurement
sample_06 Measurement
Ground truth
sample_06 Ground truth
Reference
sample_06 Reference
snapshot.npz393.7 KBmeta.json1.3 KBsolution.npz6.28 MBground_truth.npz6.22 MBground_truth.png17.1 KBsolution.png18.6 KBsnapshot.png26.4 KB
sample_07 · scene07
ref PSNR 33.8 dB
Measurement
sample_07 Measurement
Ground truth
sample_07 Ground truth
Reference
sample_07 Reference
snapshot.npz389.1 KBmeta.json1.3 KBsolution.npz6.19 MBground_truth.npz6.11 MBground_truth.png28.7 KBsolution.png29.4 KBsnapshot.png36.0 KB
sample_08 · scene08
ref PSNR 32.74 dB
Measurement
sample_08 Measurement
Ground truth
sample_08 Ground truth
Reference
sample_08 Reference
snapshot.npz392.9 KBmeta.json1.3 KBsolution.npz6.27 MBground_truth.npz6.28 MBground_truth.png18.6 KBsolution.png19.3 KBsnapshot.png25.1 KB
sample_09 · scene09
ref PSNR 34.37 dB
Measurement
sample_09 Measurement
Ground truth
sample_09 Ground truth
Reference
sample_09 Reference
snapshot.npz389.8 KBmeta.json1.3 KBsolution.npz6.24 MBground_truth.npz6.22 MBground_truth.png18.4 KBsolution.png17.9 KBsnapshot.png34.3 KB
sample_10 · scene10
ref PSNR 32.63 dB
Measurement
sample_10 Measurement
Ground truth
sample_10 Ground truth
Reference
sample_10 Reference
snapshot.npz394.0 KBmeta.json1.3 KBsolution.npz6.29 MBground_truth.npz6.17 MBground_truth.png23.2 KBsolution.png24.2 KBsnapshot.png28.6 KB
Run it
python3 scripts/regen_demos_mst_l.py  (uses pwm_core.recon.mst.mst_recon_cassi with mst_l weights)
Demo files are regenerated from scripts/generate_demos.py at fixed RNG seeds — byte-stable across runs so any SHA-256 in a sample's meta.json can be cross-checked against the repo.