A generated person can look convincing while moving impossibly because appearance and physical feasibility place different demands on a video. A planted foot must stay in contact with the floor at the right time. A hand carrying a mug must move with it. The body’s joints, support, and changing momentum have to agree throughout the action.
A realistic still leaves many of those relationships untested. Human motion exposes them. The useful question is not simply whether a clip is smooth, but whether the sequence shows a body performing the action it appears to perform.
Video models learn relationships across time
It would be misleading to describe all AI video as unrelated pictures drawn one after another. Some architectures model information across both space and time.
The Latte research paper describes a video diffusion transformer that extracts tokens spanning space and time and models video in a compressed, or latent, representation. Its variants explore ways to handle spatial and temporal dimensions. The model therefore works with relationships within a video, rather than merely generating an independent still for every instant.
That is a research architecture, not a specification of every present-day video product. Learning a distribution of plausible videos and enforcing every physical constraint are different objectives. A useful evaluation must still examine whether the generated body and its surroundings agree during the action.
Three kinds of agreement are needed
A useful way to unpack a walking figure is to separate body geometry, contact, and forces. These are linked questions, even when a review considers them one at a time.
- Body geometry: do the limbs remain connected, with joints moving within plausible limits?
- Contact: where does the body meet the world, and does that relationship stay consistent?
- Dynamics: could forces produce the observed movement?
A clip can satisfy one check while failing another. A correctly shaped leg may still slide across the floor; a stable-looking foot may accompany a body movement that its support cannot explain.
The PhyMotion preprint makes this separation operational. It reconstructs three-dimensional human trajectories from generated clips, places them in a MuJoCo humanoid model, and evaluates kinematic, contact, and dynamic feasibility. The authors report stronger agreement with human judgments than their comparison metrics in a study of 1,200 video pairs.
Those results depend on the sampled models, raters, reconstruction process, and scoring rules. A recovered body is an estimate, and simulator assumptions matter. A physics-based score can reveal errors missed by image-focused evaluation without becoming a perfect judge of all movement.
Follow a foot relative to the floor
Consider a hypothetical fixed-camera shot of someone taking one step. Choose a visible part of the supporting shoe and a floor reference beside it. While that part of the shoe is planted, it should stay approximately fixed relative to the floor. The body can move forward because the hip, knee, and ankle change together.
Once the foot lifts, movement across the floor is expected. On landing, the contact condition changes again. A reviewer who ignores those switches could wrongly flag the swinging foot for moving or overlook a supporting foot that quietly skates.
This suggested inspection has limits. A moving camera shifts image coordinates even when contact is sound. Occlusion can hide the shoe. A featureless floor may provide no reliable reference. In those cases, record that the relationship is difficult to judge rather than declaring a physical violation from the screen position alone.
Smoothness cannot establish support
Imagine an animation where a shoe glides evenly across a level floor while the person seems to place weight on it. Removing jitter could make the sliding smoother while leaving the contact problem intact. Conversely, a genuine footfall can include a quick change in motion.
This is why “make it smoother” is an incomplete diagnosis. Describe what failed: contact began too late, a supporting foot drifted, or a hand moved through the object it should have held. A precise observation gives a revision something concrete to address.
The PhysDiff research paper tackles floating, foot sliding, and ground penetration in generated human motion. It uses motion imitation in a physics simulator during diffusion to guide motion toward a physically plausible trajectory. Its output concerns human-motion data, rather than solving every problem in finished photorealistic footage. It illustrates how explicit constraints can participate in generation itself.
Guide the action before polishing the surface
Brown University’s GenHSI project takes a different route. It breaks a description into actions, constructs three-dimensional keyframes with contact cues, and uses them to guide video generation involving a person and a scene. Planning where the body interacts with furniture helps specify more than the final image should look like.
The method uses particular representations and pretrained systems. Its demonstrations do not imply that any arbitrary interaction can now be generated correctly. They show a concrete way to supply structure that a short text description may leave implicit.
Another research result challenges the assumption that refinement always improves motion. The 2026 PhaseLock project reports cases where two-step generation preserved physical consistency better than fifty-step generation from the same model. Its method carries early motion information into later visual refinement. This is a finding under the paper’s tested conditions, not a general instruction to minimize generation steps in every tool.
Review the action the clip actually needs
For an original hypothetical clip of a person standing from a chair and lifting a mug, a short review could ask:
| Moment | Relationship to inspect | If it is hidden |
|---|---|---|
| Rising | Does the body leave the seat coherently? | Do not infer the full movement from the final stance |
| Reaching | Can the arm approach without passing through the table? | Mark the approach as unverified |
| Gripping | Do the fingers and mug meet before moving together? | Check what contact is actually visible |
| Carrying | Does the mug remain coupled to the hand? | Inspect reappearance without assuming continuity |
These are proposed review questions, not reported benchmark results. For a broader method covering object identity and scene changes, use the AI video continuity check.
An awkward movement alone does not prove that footage is generated, and a convincing movement does not authenticate it. The narrower judgment is still useful: decide whether this clip visibly performs the required action. Attractive frames can establish the look. The contacts and transitions establish whether the action holds together.
