When a storm can head north or south, a forecast down the middle may score better under squared error than either plausible path. The same paradox appears when video generators are tested as world models: fewer denoising steps can raise physics scores even as videos blur or stop moving early. Why? Physics-IQ compares each generated video with one recording, although a scene can have multiple valid futures. An exact decomposition shows its scores can reward reduced variation among generated futures, even when their distribution becomes less accurate. We introduce Kernel Physics-IQ to remove this incentive in expectation, needing only two independent samples per condition. Two-step sampling lowers Kernel Physics-IQ in all five image-to-video models. For CogVideoX, the official score rises by 0.055 while ours falls by 0.131. Changing the sampling schedule reveals further differences official scores miss. In controlled physical systems with known outcome probabilities, our metric favors the true distribution over collapsed predictions. In short, a world model should be judged not only by how closely one video matches what happened, but by how faithfully its predictions represent what could have happened.
Figures & tables
Figure 2: Two-step samples blur, agree across seeds and freeze early. Wan2.2 on silk-cover (a) and rolling-reflection (b), scenes whose 50-step samples keep objects intact: each frame carries its 0–5 s motion mask, colored by when each pixel last moves; ticks mark when each take or seed stops. In silk-cover (takes settled by 2.5 s), all 50-step seeds still move at 5 s; two-step seeds settle on one blurred teapot by 4 s. In rolling-reflection (takes stop by 2.9 s), two-step seeds smear the ball and stop by 1.75 s; 50-step seeds roll until 2.75–5 s.
Figure 3: What one sample cannot see. A sample is scored by its distance to the recording (blue), which on average is the kernel score plus half the self-dispersion (green). Two-step samples disperse less: the dispersion premium they collect (dashed) outweighs their kernel loss, so single-sample scoring prefers them. A second sample measures the self-dispersion, which Kernel Physics-IQ subtracts. Frames: paper-smoke 2 s after the switch frame (Wan2.2, two seeds). Brackets (root-mean-square) and bars: Wan2.2 ST IoU distances on whole clips, mean over 66 scenes.
Setting
Official ↑
Kernel Physics-IQ ↑
SSR
CogVideoX, 50 steps
0.301
0.737
0.90
CogVideoX, 2 steps
0.356
0.607
0.29
CogVideoX, LDG (50 steps)
0.351
0.707
0.81
Cosmos, 35 steps
0.294
0.725
1.09
Cosmos, 2 steps
0.301
0.609
0.51
Wan2.1, 50 steps
0.291
0.746
1.21
Table 1: Two steps lower Kernel Physics-IQ for every model. Center-view scores on all 66 scenes (SSR is 1 when calibrated). Bold marks the higher point estimate between default full and two-step runs. HunyuanVideo’s SSR falls via sample-mean error without detected net shrinkage. Take 2 is scored as a one-outcome forecast on the 52 non-duplicate pairs.
Figure 4: Fewer steps align seeds far more than they match truth. (a) Recording ( xˉ ) vs. seed overlap ( yˉ ) with Eq. 6 contours and the 0.695 no-overlap line. (b) Dispersion-premium and kernel-part bars behind net diamonds for the official split. (c) Spatial masks behind x and y : Wan2.2 seeds 43 and 44 on rolling-reflection ( Fig. 2 b) against take 1 (teal). (d) ST IoU splits by scene motion.
Figure 5: Fixed-weight interventions expose the dispersion premium. (a) Sweeping CogVideoX node t2 lifts Kernel Physics-IQ and SSR past 499 with no detected official difference. (b, c) CogVideoX and Cosmos steps. (d) Cosmos history arms ( n=32 , n=15 shuffled).
Figure 6: Tied official scores hide different futures. (a) Predictors in a pendulum world that swings left or right with equal probability: each card shows what the predictor draws (line weight: probability; dashed: never drawn), and the pixel mean blends both swings into one video (diamond: mean position). (b) Exact scores, averaged with a ball dropped onto a peg: the raw official composite ties the first four, whereas Kernel Physics-IQ ranks the true distribution first, as it does among all 34 predictors. (c) Mask distance: staying still wins below the thresholds of Prop. 2 .