Physics in 2-Steps: Why 2 Denoising Steps Beat 50 for Physical Consistency in Video Diffusion

Forum topic · 小凯 · 2026-06-12

Summary

Researchers from Yonsei University and NVIDIA discovered a counterintuitive finding in image-to-video (I2V) diffusion models: running only 2 denoising steps produces videos with better physical consistency than 50 steps, despite lower visual quality. Through frequency analysis, they show the denoising chain preserves the Fourier magnitude spectrum (texture, appearance) but progressively degrades the phase spectrum (motion, structure), with phase corruption causing 8.5x more optical flow distortion than magnitude corruption. They propose PhaseLock, a training-free, model-agnostic framework: extract a motion prior (inter-frame latent delta) from a 2-step run, then use it as Latent Delta Guidance during a full 50-step denoising run. PhaseLock improves Physics-IQ by +6.2 on average across CogVideoX and Wan 2.1 with only 1.06x time and 1.02x memory overhead, while preserving visual fidelity. It is roughly 5x more efficient than external guidance methods like reward models or physics simulators, requiring no extra training, data, or tools.

Physics in 2-Steps: Locking Motion Priors Before Visual Refinement Erases Them

A surprising discovery from Yonsei University and NVIDIA (arXiv:2606.06361, ICML 2026): increasing denoising steps from 2 to 50 in I2V diffusion models makes videos *less* physically realistic — not in visual quality, but in physics. The denoising chain refines textures while quietly erasing correct motion priors. The root cause lies in the Fourier phase spectrum.

The Counterintuitive Finding

I2V diffusion models produce visually stunning videos, but motion often violates physics: balls pass through walls, stick to surfaces, or bounce with impossible energy.

Existing fixes are expensive:

External physics simulators (Blender, PyBullet) guiding generation — heavy overhead
Reinforcement learning / reward models (e.g., WMReward) — extra data and training required

The surprising result: the same model with the same input produces more physically consistent video at 2 denoising steps than at 50. The 2-step output is noisier and blurrier, but its motion trajectories are correct.

Root Cause: Phase Spectrum Degradation

Spatiotemporal Fourier analysis shows information splits across:

Magnitude spectrum — texture, appearance, detail
Phase spectrum — structure, motion trajectories, object positions

Key finding: denoising mainly destroys phase while magnitude stays relatively stable (phase drops ~18% from 2 to 50 steps).

Sensitivity experiments confirm this: phase corruption causes 8.5x more optical flow distortion than magnitude corruption. Physical motion is highly phase-sensitive; denoising is "preserve magnitude, destroy phase."

A control experiment rules out blur as the explanation: Gaussian-blurring the 50-step output to match the 2-step blur level still leaves the 2-step output with significantly higher Phase Temporal Correlation. The phase loss is structural, not a frequency artifact.

PhaseLock: A Two-Step Strategy

PhaseLock is training-free and model-agnostic.

Step 1 — Extract the motion prior (2 steps):

Run 2 denoising steps to get a physically-correct-but-coarse latent sequence $$z_2$$
Compute the inter-frame motion delta: $\Delta_{phys} = z_2(t+1) - z_2(t)$

Step 2 — Lock and guide (50 steps + guidance):

Restart the full 50-step denoising run
Apply Latent Delta Guidance at each step:

z_{t+1} = z_{t+1}^{denoise} + \lambda \cdot \Delta_{phys}

with $\lambda = 0.05$ , linearly decayed

Working in latent space (rather than pixel space) preserves denoising dynamics and avoids damaging texture details.

Results

Tested on CogVideoX and Wan 2.1 (plug-and-play for latent-based diffusion models):

| Metric | Baseline (50 steps) | +PhaseLock | |--------|--------------------|------------| | Physics-IQ | baseline | +6.2 (average) | | Visual fidelity | high | high (preserved) | | Time overhead | 1x | 1.06x | | Memory overhead | 1x | 1.02x |

Qualitative improvements: fluid poured from a cup follows gravity instead of passing through walls; balls rolling off tables follow Newton's laws instead of floating or bouncing forever; capillary flow respects surface tension. A striking example: a magnet basket lifted over a tennis ball — the baseline makes the non-magnetic ball stick to the magnet; PhaseLock leaves it on the table.

Efficiency comparison: PhaseLock runs at 1.06x with no training or external tools, versus ~5x for reward-model guidance (WMReward) and >5x plus a physics engine for simulator-guided methods — roughly 5x more efficient than comparable external-guidance approaches.

Ablations: $\lambda = 0.05$ is optimal; 2 steps is the best prior-extraction point (1 step is unstable, >2 steps phase starts degrading); latent-space guidance clearly beats pixel-space.

Why It's Elegant

PhaseLock requires no training, no reward models, no physics engines, no fine-tuning, no physical annotations. The deep insight: the model already "knows" correct physics at 2 steps — 50 steps doesn't fail to learn physics, it forgets physics it already had. Diffusion models are trained to make videos look real, not to obey physical laws; when the optimization target is visual fidelity alone, physical correctness becomes an expendable variable.

Limitations

Best suited to scenes with clear physical motion (rigid bodies, fluids, collisions) and latent-based diffusion models
The 2-step prior is relatively better, not absolutely perfect — complex multi-object or soft-body physics may still be beyond it
Complementary to external guidance methods like WMReward, not a replacement

References

Han et al., "Physics in 2-Steps: Locking Motion Priors Before Visual Refinement Erases Them", arXiv:2606.06361 (ICML 2026)
Institutions: Yonsei University, NVIDIA
Project page: https://dnwjddl.github.io/phaselock/