Physics in 2-Steps: Why 2 Diffusion Steps Understand Physics Better Than 50 — PhaseLock Explained

> A finding that chills diffusion model researchers: increasing denoising steps from 2 to 50 makes videos *less* physically realistic — not in image quality, but in physics. A team from Yonsei University and NVIDIA found that the denoising process quietly "erases" correct motion priors while refining textures. The cause hides in the Fourier phase spectrum.

---

1. A Counterintuitive Finding: Less Is More

Image-to-video (I2V) diffusion models produce stunning visuals, but a long-overlooked problem remains: videos look beautiful, yet their motion often violates physics.

A basketball hitting a wall should bounce. Instead, generated videos show the ball passing through the wall, sticking to it, or bouncing forever in violation of energy conservation.

Existing solutions take two routes: 1. External physics simulators: use traditional engines (Blender, PyBullet) to generate reference motion. Works, but expensive. 2. RL / reward models: train a physics-consistency reward model (e.g., WMReward) and optimize via RL or classifier-free guidance. Requires extra data and training.

This paper (arXiv:2606.06361, ICML 2026) found something completely unexpected:

> The same model, same inputs — video generated with only 2 denoising steps has better physical consistency than with 50 steps.

This is not an image-quality-vs-physics tradeoff. The 2-step output is indeed rough (noisy, blurry textures), but its motion trajectories are correct. At 50 steps, quality is maxed out while motion goes wrong.

---

2. Root Cause: Phase Spectrum Degradation

2.1 Fourier View: Motion and Texture Live in Different "Bands"

Applying a spatiotemporal Fourier transform to video:

• Magnitude spectrum: encodes texture, appearance, details
• Phase spectrum: encodes structure, motion trajectories, object positions

Key finding: the denoising process mainly destroys the phase, while magnitude is barely affected.

| Metric | 2 steps → 50 steps | |------|---------------| | Phase spectrum | ~18% drop | | Magnitude spectrum | Relatively stable | | Physical consistency | Significantly worse |

In other words, the denoising chain optimizes magnitude (making images look sharper and more realistic) at the cost of progressively corroding the phase — exactly where physical motion is encoded.

2.2 Phase Sensitivity

Corruption experiments:

• Phase corruption: add noise to the phase spectrum, keep magnitude
• Magnitude corruption: add noise to the magnitude spectrum, keep phase

Result: phase corruption distorts optical flow 8.5x more than magnitude corruption. Physical motion is highly sensitive to phase; visual appearance to magnitude. Denoising happens to be "preserve magnitude, destroy phase" — so the more it denoises, the worse the physics.

2.3 Control Experiment: Not a Blurriness Artifact

Could 2-step videos merely look more "physical" because blur hides errors? The paper rules this out: after matching blurriness of the 50-step output to the 2-step level via Gaussian blur, 2-step outputs still show significantly higher Phase Temporal Correlation.

> Phase loss is structural, not a frequency artifact.

---

3. PhaseLock: A Two-Step Strategy

PhaseLock is a training-free, model-agnostic framework.

3.1 Core Idea

50 steps are not needed to establish physics — 2 steps already capture the correct motion prior. The question is how to preserve it during subsequent visual refinement.

Step 1: Extract the motion prior (2 steps)

• Run 2 denoising steps to get a rough but physically correct latent sequence \(z_2\)
• Compute inter-frame motion deltas: \(\Delta_{phys} = z_2(t+1) - z_2(t)\)
• This delta encodes physically valid motion evolution

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

• Restart the 50-step denoising
• At every step, add Latent Delta Guidance:
\[z_{t+1} = z_{t+1}^{denoise} + \lambda \cdot \Delta_{phys}\]
• With \(\lambda = 0.05\), linearly decayed
• This biases the 50-step trajectory to stay near the physically correct path already found

3.2 Why Latent Delta?

Pixel-space guidance would destroy texture detail. Operating in latent space:

• Preserves the model's denoising dynamics
• The delta encodes motion, minimally affecting texture
• Integrates seamlessly with any latent-based diffusion model

---

4. Experiments: Plugs into Mainstream Models

4.1 Tested Models

| Model | Type | Integration | |------|------|------| | CogVideoX | I2V diffusion | Direct | | Wan 2.1 | I2V diffusion | Direct | | Other latent-based models | Latent diffusion | Plug-and-play |

4.2 Core Results

| Metric | Baseline (50 steps) | +PhaseLock | Gain | |------|------------|-----------|------| | Physics-IQ | baseline | – | +6.2 (avg) | | Physical consistency | Low | High | Significant | | Visual fidelity | High | High | No loss | | Time cost | 1× | 1.06× | Negligible | | Memory cost | 1× | 1.02× | Negligible |

Key: PhaseLock improves physics without sacrificing visual quality — it guides refinement along a physically correct direction rather than constraining the model into coarseness.

4.3 Qualitative Comparisons

Scenarios include:

• Fluid displacement: baseline water pouring may pass through walls; PhaseLock obeys gravity
• Rigid-body drops: baseline balls may float or bounce forever; PhaseLock follows Newton's laws
• Capillary flow: baseline may violate surface tension; PhaseLock stays consistent

A particularly intuitive demo (project page):

> Magnetic basket experiment: a basket with a magnet is lowered over an orange tennis ball, then lifted. > - Baseline: the ball sticks to the magnet and gets lifted (tennis balls are not magnetic!) > - PhaseLock: the ball correctly stays on the table

4.4 Efficiency Comparison

| Method | Time | Training required | External simulator | |------|---------|------------|------| | Baseline (50 steps) | 1× | No | No | | PhaseLock | 1.06× | No | No | | WMReward (external guidance) | ~5× | Reward model | No | | Simulator-guided | >5× | No | Yes |

PhaseLock is ~5x more efficient than comparable external-guidance methods, with no extra training or tools.

4.5 Ablations

• Lambda: 0.05 is optimal — too small gives weak guidance; too large harms texture
• Step count: 2 steps is the best prior-extraction point; 1 step is unstable, >2 steps phase already degrades
• Latent vs pixel space: latent-space guidance clearly better for texture preservation

---

5. Why the Method Is Elegant

| Dimension | Traditional methods | PhaseLock | |------|---------|------| | Training | Reward models / RL | None | | External tools | Physics engines | None | | Model changes | Fine-tuning | None | | Data annotation | Physics labels | None |

The deep insight: the I2V model already "knows" correct physics at 2 steps — it just lacks steps to polish textures. The problem at 50 steps is not "failing to learn physics" but "forgetting physics it already knew."

Like a student:

• 2 steps = quickly drafting the correct answer (right idea, messy handwriting)
• 50 steps = spending all the time on neat handwriting while corrupting the formula
• PhaseLock = lock in the draft's correct reasoning, then polish the handwriting

---

6. Limitations and Discussion

6.1 Scope

Best suited for:

• Scenes with clear physical motion (rigid bodies, fluids, collisions)
• Latent-based diffusion models (CogVideoX, Wan 2.1, etc.)

Possibly unsuitable for:

• Highly stylized / non-physical motion
• Non-latent diffusion models
• Scenarios requiring precise physical parameters (engineering simulation)

6.2 The 2-Step Prior Is Not Perfect

The authors acknowledge that the 2-step motion prior, while more physical than 50 steps, is still imperfect. PhaseLock provides *relative* improvement, not an absolute solution — complex multi-object interactions or soft-body physics may exceed the prior.

6.3 Relation to External Guidance

PhaseLock does not replace WMReward-style external guidance; it offers a much lighter baseline that can be combined with external guidance if needed.

---

7. Conclusion: The Other Side of Denoising

Denoising is designed to "generate better-looking images." But this paper reveals a neglected dimension: while optimizing visual fidelity, denoising corrodes the phase information encoding physical motion in latent space.

This is not a bug — it's a side effect of the training objective. Models are trained to make videos that *look* real, not videos that *obey physics*. When only visual quality is optimized, physical correctness becomes the sacrificial variable.

PhaseLock's elegance: it doesn't modify the model or add external constraints, but exploits the model's "knowledge gap" across denoising stages — physics knowledge is present at 2 steps, texture knowledge at 50. Use the former to guide the latter.

> "PhaseLock extracts Δphys from 2 steps and injects it back into the 50-step run, recovering both fidelity and physics."

1.06x time cost for +6.2 Physics-IQ — possibly one of the best trade-offs in video diffusion.

---

References

• Paper: 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/
• Key results: +6.2 Physics-IQ, 1.06× time, 1.02× memory
• Compatible models: CogVideoX, Wan 2.1, etc.