Rather
World model benchmarks
Physics-IQ
Physical prediction in generated video across mechanics, fluids, optics, thermodynamics and magnetism. Original and Verified results use separate protocols.
Published resultsBenchmark website (opens in a new tab)
Do both objects fall naturally?
Physics-IQRather-generated AI video · Preview · Unscored
Do both objects fall naturally?
0:00 / 0:05
About this comparison
What to look for Compare the release, fall and contact with the pillows. The reference video shows what actually happened.
Starting frame ↗The task & prompt
Generate the scene described below.
Two pillows on a table and two grabber tools hanging above them from which a brown tennis ball and an orange block are suspended. The grabber tools let go of the ball and block. Static shot with no camera movement.
Physics-IQ Original / I2V / original prompt / case 0001
Sources & generation details
Generated by Rather. An official benchmark input run with preview settings. No benchmark scores have been computed for these clips.
Case and seed 42 were fixed before generation. All 2 planned model outputs are shown; no best-of selection.
These exact checkpoints were run by Rather. They are not linked to similarly named source-reported leaderboard entries. Full prompts, model revisions, settings and output hashes are in each run record.
768 x 448 visual preview; not a full benchmark evaluation. One preselected seed per model; no output selection or prompt optimization. Original dataset and original prompt, not Physics-IQ Verified. Reference is not exposed to generation. Generated video is approximately five seconds at 24 fps; no Physics-IQ score has been computed. Conditioning images are resized to the full output frame without cropping. Generated MP4 streams are remuxed for playback; VP9 WebM copies are encoded at CRF 24 for browser compatibility with no frame interpolation. Original outputs remain linked. Reference footage, where provided, is resized to 768 pixels wide; source hashes and attribution are preserved. Input licenses are listed here; model licenses are linked in each run record.
Copyright 2025 Google LLC. Physics-IQ authors. Model generations by Rather.
Wan 2.2 TI2V 5B · Seed 42 · 768 × 448 · 121 frames per turn · 24 fps · 40 steps · Guidance 5. Full run record & prompts (opens in a new tab)
LTX-Video 0.9.5 2B · Seed 42 · 768 × 448 · 121 frames per turn · 24 fps · 40 steps · Guidance 5. Full run record & prompts (opens in a new tab)
Benchmark source (opens in a new tab)CC BY 4.0 license (opens in a new tab)Wan 2.2 TI2V 5B original clip (opens in a new tab)LTX-Video 0.9.5 2B original clip (opens in a new tab)
Source revision: c2bca8a5bf452d24e177e9975641af1888eb1492
Link to this example ↗13 reported results
% · Higher is better · Published by Physics-IQ
| Compare | Model / variant | % | Evidence |
|---|---|---|---|
Odyssey-3 Pro+ best-of-8 (WMReward + MBR consensus, rank-sum) (custom prompts) / multiframe (v2v)Not specified by source | 66.10 | View source ↗Evaluator unspecified | |
Odyssey-3+ best-of-8 (WMReward + MBR consensus, rank-sum) (custom prompts) / multiframe (v2v)Not specified by source | 64.43 | View source ↗Evaluator unspecified | |
FLUX 3 [large](Claude Opus 5.5 prompts) + WMReward + consensus (BoN) / multiframe (v2v)Not specified by source | 64.35 | View source ↗Evaluator unspecified | |
Odyssey-3 Pro(custom prompts) / multiframe (v2v)Not specified by source | 63.37 | View source ↗Evaluator unspecified | |
Odyssey-3(custom prompts) / multiframe (v2v)Not specified by source | 61.56 | View source ↗Evaluator unspecified | |
FLUX 3 [large](Claude Opus 5.5 prompts) / multiframe (v2v)Not specified by source | 61.11 | View source ↗Evaluator unspecified | |
Magi-1 24B+ GeoPhys (BoN) (op) / multiframe (v2v)Not specified by source | 58.20 | View source ↗Evaluator unspecified | |
Physical Registry + PRS (BoN=8, Cosmos 3 Nano 16B)(bpp) / multiframe (v2v)Not specified by source | 51.86 | View source ↗Evaluator unspecified | |
Odyssey-3(bpp) / multiframe (v2v)Not specified by source | 51.76 | View source ↗Evaluator unspecified | |
Cosmos3-Supermultiframe (v2v)Not specified by source | 50.80 | View source ↗Evaluator unspecified | |
Physical Registry (Cosmos 3 Nano 16B)(bpp) / multiframe (v2v)Not specified by source | 50.17 | View source ↗Evaluator unspecified | |
Magi-1 24B(op) / multiframe (v2v)Not specified by source | 48.40 | View source ↗Evaluator unspecified | |
Cosmos3-Nanomultiframe (v2v)Not specified by source | 43.00 | View source ↗Evaluator unspecified |
Physics-IQ Verified with multiframe video conditioning. Revised prompts and scoring; not interchangeable with original Physics-IQ. Methods may use custom prompts and best-of-N selection; see evidence. A reported score is not an independent verification by Rather.