Image-to-Video diffusion models leverage input images to generate visually stunning content, yet frequently produce motion that violates physical laws. We reveal a surprising finding: a 2-step generation often exhibits better physical consistency than a 50-step output from the same model. Through spectral analysis, we trace this to phase erosion during denoising; the phase degrades significantly (dropping by $\approx 18\%$ from step 2 to step 50), whereas the magnitude remains relatively stable. Building on this insight, we propose PhaseLock, a training-free framework that preserves the valid motion priors from few-step inference throughout the denoising trajectory. Rather than relying on full-step inference for physical consistency, PhaseLock extracts a motion prior from just 2 steps and enforces it onto high-fidelity generation via Latent Delta Guidance. Our approach effectively mitigates phase degradation, improving physical consistency by an average of 6.2 points across diverse models while largely maintaining visual fidelity, with negligible overhead ($1.06\times$ time, $1.02\times$ memory) and reduced reliance on expensive external guidance methods ($\sim5\times$ time).
Alex: Welcome to another episode of ResearchPod.
Sam: Today we're looking at a genuine paradox in AI video generation. These systems can produce footage that looks completely realistic, yet they often fail at something a child understands instinctively—like a ball falling in a straight line, or water flowing downhill. And the central finding is striking: these models actually work out the correct physics early in the generation process, but then erase that knowledge while polishing the final visual details.
Alex: So the model isn't missing the knowledge—it's forgetting what it already figured out?
Sam: That's the key insight. Think about how these AI video systems work. They don't draw a video from scratch. Instead, they start with pure visual noise—like static on an old TV—and gradually clean it up over many steps. That cleaning process is called denoising. And researchers found something telling: a video cleaned up in just two steps often follows physical laws more faithfully than one cleaned up in fifty.
Alex: That seems backwards. More steps should mean better results, right?
Sam: You'd think so. But here's what's actually happening. In those first couple of steps, the model establishes a rough motion plan—where things are going, how fast, in what direction. It's like roughing out a sketch before you paint. The problem is that as the model spends the next forty-eight steps adding fine visual detail, it gradually overwrites that motion plan. The painting ends up beautiful, but the movement no longer makes physical sense.
Alex: So how do the researchers actually measure this motion plan being lost?
Sam: They describe it using something called the phase spectrum. Here's the intuition: any image or video can be thought of as built from waves of light and color layered on top of each other. Each wave has two properties. The first is its strength—how bright or dark it makes things. The second is its timing—where exactly things are positioned and how they move from frame to frame. Researchers call that timing component the phase. What they found is that the strength stays relatively stable throughout the fifty-step process, but the phase—the part that encodes motion—degrades significantly.
Alex: So the phase is essentially the blueprint for movement. Corrupt that, and the physics falls apart.
Sam: Exactly. And they tested this directly. They took real videos and deliberately corrupted either the phase or the strength. When they scrambled the phase, motion became erratic and physically implausible. When they scrambled the strength, the video barely changed in terms of how things moved. Motion is almost entirely encoded in the phase, and the phase is what the denoising process slowly destroys.
Alex: So they've diagnosed the problem. What's the fix?
Sam: They built a method called PhaseLock. The idea is straightforward: run the first two steps of the generation process, extract the motion blueprint from the phase at that point, and then use that blueprint to guide the remaining forty-eight steps. You can think of the guiding mechanism as a fence. As the model refines the visual details, the fence keeps nudging it back toward the motion plan it established at the start.
Alex: Like keeping a rough sketch pinned to the wall while you paint, so you don't lose track of where everything was supposed to go.
Sam: That's a good way to put it. And importantly, this is what researchers call a training-free approach—they don't have to rebuild or retrain the AI model from scratch. PhaseLock works on top of existing models, which makes it practical to apply broadly. It also adds minimal extra computation time, so it avoids the heavy cost of bolting on a separate physics simulation engine.
Alex: So in a sense, this is about keeping the AI honest. It already understood the physics—the fix just stops it from talking itself out of that understanding.
Sam: That's the core of it. The physics knowledge was always there. The model just needed a mechanism to preserve it through the refinement process, rather than letting it get smoothed away in the pursuit of visual polish.
Alex: It's a meaningful reminder that more processing doesn't always mean better reasoning—sometimes the clearest thinking happens early, and the challenge is holding onto it. Thanks for listening to ResearchPod.