How do I check a generated video for physics failures before I publish it?

October 23, 2026 · 4 min read

SentX SI

Before a generated clip goes anywhere public, give it a short physics pass: write down what should happen before you watch, view the clip in slow windows, log every moment where what happens conflicts with what should, and fix only the parts that truly break. It sounds fussy for a five-second clip, but the failures that make viewers drop a video — an object slipping through another, a ball bouncing the wrong way, a hand merging into a table — are usually visible to someone looking for them, and they are cheap to catch before upload and expensive after.

The structure below adapts a paper submitted to arXiv on October 22, 2026, called PhysTrace, by Taekoan Yoo and three co-authors. The paper attacks the same problem from the machine side: it compares what a multimodal model expects to happen next, given the frames so far, against what actually appears in the following window, and flags the interval where the two disagree. Two of its ideas transfer directly to a human reviewer. Violations are events with a location in time, so the useful output of a review is a timestamped list, not a vague sense that something is off. And spotting a suspicious moment is not the same as confirming a violation — the paper separates possible from impossible, and a reviewer has to make that same call. To be clear about what this is: the checklist is an editorial adaptation of that framing for creators, not a result the paper measured.

Source: PhysTrace: Interpretable Temporal Evaluation of Physical Consistency in Video Generative Models

Have your prompt or scene description open while you review, and know the format you will publish in — if the clip is going out vertical, check it vertical. Then work through the pass.

Source: AI Video Generator — Text & Image to Video | SentX SI

1. Write down what should happen. From your prompt and from ordinary physics: what moves, what touches what, what falls, what the camera does. A line or two per element is enough. This is the expectation set you will test the clip against.
2. Watch once at normal speed, without pausing. You are checking whether the story reads in one go — does the subject do what you asked, does the motion land. Note anything that catches your eye with its timestamp, but do not stop to investigate yet.
3. Do a slow pass. Drop the playback speed or step frame by frame, and spend your time on the moments where things meet or change state: contacts, impacts, starts and stops of motion, anything crossing the frame edge. Most clipping, momentum, and gravity errors live in these windows, and they survive a fast watch because they last only a few frames.
4. Log each conflict as time plus description. "0:04–0:06, cup slides through the table edge" is actionable; "the middle part looks weird" is not. If you cannot pin a moment to a timestamp, loop it until you can.
5. Sort the log into three buckets: true violations, ambiguities, and intended style. Slightly shifting character proportions might be your aesthetic; a coffee cup passing solidly through wood is not. Only the first bucket blocks publication.
6. Fix one variable at a time. Change the phrase describing the failing action, or swap the starting image, and regenerate. Re-run the pass on the new clip, focusing on the segment that failed before. Change three things at once and you cannot tell what fixed what.Source: AI Video Generator — Text & Image to Video | SentX SI
7. Check the final cut at posting size. On the phone, at the ratio you will publish in: does the motion still read, does the composition clear the area where the platform puts its buttons and captions?

The categories in step five are the ones reviewers most often run into when checking generated clips: objects occupying the same space, gravity acting in the wrong direction or with the wrong strength, motion that stops or continues without cause, things appearing or vanishing mid-scene, feet sliding on ground, shadows and light that do not track their sources, and flicker or drift that creeps in over the clip's length.

The honest boundary: the underlying research evaluated benchmark clips, and its companion dataset, IntPhys2-Timing, marks the exact intervals where violations occur in those scenes. Even there, the system's own reported hit rate tops out at 70.7 percent across physical principles — meaning a purpose-built evaluator still misses roughly three out of ten violation windows, and that figure is the authors' own, on their own benchmark. Complex multi-object scenes, long takes, and subtle interaction errors are where both machines and humans get weaker. The checklist narrows your miss rate; it does not eliminate it.

Source: PhysTrace: Interpretable Temporal Evaluation of Physical Consistency in Video Generative Models

If you are working in SentX's video generator, this is the review step its own instructions point at. The generator page tells you to review the motion and visual details before use, and its refine loop — adjust the wording or the starting image, regenerate until the motion feels right — is built for exactly this workflow. Ten minutes of structured watching is the difference between a clip that holds up and one that ends up in a screenshot thread.

Source: AI Video Generator — Text & Image to Video | SentX SI; Victoria AI Model Card | Capabilities, Evidence and Limits | SentX SI