10 Motions That Break a 5-Second AI Video — Why We Banned Them and What We Use Instead

Banned motions stay defined in our library but can never be selected. This article covers 10 of them:

The reasons collapse into two groups:

• Too fast for five seconds — the spins, the hop and the choreography (6)
• The gesture doesn't read — viewers can't tell what it means (4)

You can't talk about motion failures without the pipeline, so here it is:

1. Generate one start image with SDXL plus PuLID, which locks the character's facial identity
2. Hand that image to Wan 2.2, an image-to-video model, along with a motion instruction written in English
3. The model renders 81 frames at 16 fps — about 5.06 seconds
4. A human reviews every clip before it is approved or rejected

The motion instruction is assembled from a library of predefined motions: one or two of them (usually two), joined with `, then ` into a single sentence, for example:

she lifts one hand and gives a small playful wave to camera, ... , then her gaze drifts off to the side as if thinking about something, ...

Our first suspect was simply how much motion we were cramming into five seconds. We counted the then separators in each clip's motion instruction — each one is a change of action, a "beat" — and compared that against how often human review rejected the clip.

Rejection rate rose steadily with every added beat.

The arithmetic explains why. Chain two motions into 5.06 seconds, give each an internal "spin → stop → pose" structure, and you get five beats in total — roughly one second per action. That's the speed at which a clip starts to look frantic.

Here are the six we removed. Every one of them sounds charming on paper.

We recorded two reasons:

1. Too fast for five seconds. As the numbers above show, each action gets about a second
2. Spinning takes the face off camera, and the facial identity fixed in the start image can't be held once it comes back

The second reason follows directly from the pipeline: the model only has the start image, so the face that completes the turn is a face it drew from scratch.

If too many beats is the problem, cap the beats. So we set a hard limit: at most 3 beat changes per clip. Then we found something we hadn't expected:

• 5 of the 6 banned motions contained no `then` at all
• "Spin, hair flares, land facing camera, laugh" had been written as one comma-separated action

By the counter's logic, these were zero-beat, calm motions.

The cap actually changed only 1.0% of clips, and the average barely moved.

These three were removed for meaning, not speed.

The verdict in review: you can't tell whether she's surprised or laughing.

In a still photo, a hand over a smiling mouth reads naturally. In five seconds of video, the moment the hand comes up could be surprise, laughter or embarrassment — and the clip ends before it resolves into any of them.

This is often misread, so: hands near the face are not banned. Motions still in use include:

• Blowing a kiss — kept deliberately, because its intent is unambiguous and the ban's reason doesn't apply
• Peace signs and finger hearts
• Covering the face with both hands and peeking through the fingers
• Resting a hand on the cheek with a head tilt

Two days after the hand-to-mouth ban, one more motion went for the same reason.

The recorded reason matches the first group: in a five-second clip, a finger brought to the mouth becomes a picture of someone doing something unclear. A "thinking" pose that works in a photograph simply didn't read as thinking in motion.

In its place we use a sideways glance that drifts back slowly, and a dreamy chin-on-hand pose.

Some failures didn't need a motion removed — just a different way of writing the instruction. If you generate your own clips, these may be the most useful part of this page.

Posture conflicts turned up when we machine-generated 3,000 motion specifications as a test: 11.0% (330) contained incompatible postures. We fixed the selection step to align postures — and 6 (0.2%) still slipped through.

The cause: a later stage recombined the selected motions with results from a separate draw. Every part was correct; the assembly broke it.

Separate from the ban list, a recount of the motion library turned up a different problem.

As of August 2026, 111 of 246 motions (45%) involved smiling or laughing. Among the general-purpose motions we use most, it was 33 of 37 (89%). Wistful expressions? Just 2.

No single clip had anything wrong with it. But lined up in a feed, every character was smiling the same smile. Individual videos fine; the site as a whole, monotonous.

The fix wasn't rewriting existing motions. It was adding options that don't smile:

• One slow breath, shoulders settle, a calm steady look at the camera
• A gaze that drifts aside as if thinking, then returns unhurried
• Eyes closed for a beat as the air moves her hair
• A small head tilt, as if listening to something off camera

Here is everything above, condensed into a checklist for image-to-video generation. We verified it on 5-second Wan 2.2 clips, but the reasoning is about clip length and human viewers, so it should carry over to other short-form models.

1. Three beat changes at most in five seconds — one or two is better
2. Count comma-joined actions too. "Spins, stops, laughs" is three actions with no then in sight
3. Avoid full-body spins, jumps and fast shaking. Turn the head or shoulders instead
4. Keep the face in roughly the start image's orientation. A face that leaves the frame gets redrawn
5. Bring hands to the face only for gestures with one clear meaning — a wave, a peace sign, hiding behind both hands. Skip hand-to-mouth and finger-to-lip
6. Don't write "repeats" or "several times"
7. Don't write negations. Describe the opposite state in positive terms
8. Don't ask for color to change during the clip
9. When chaining two motions, keep the posture the same
10. Keep instructions short. We cut the median video-stage prompt from 799 to 576 characters

None of the 10 banned motions in this article were deleted. Their definitions stay in the library; the selection step simply rejects them every time.

Three reasons for doing it this way:

• Re-enabling is a one-line change. If a future model renders spins cleanly, we can re-test and restore
• The reason for each ban sits right next to the motion in the code
• When and why each ban happened stays traceable

Every number in this article was recounted from that record. The dates are there too: the three hand-to-mouth motions, the six spin/hop/choreography motions and the finger-on-lip motion were added on three consecutive days in late August 2026.