A Motion-Continuity Test for AI Video Generators #

A polished first frame does not prove that an AI video generator can maintain a usable shot. Production failures often appear only after motion begins: identity drifts, hands merge with props, camera direction changes without cause, and background geometry bends between frames.

This protocol measures continuity with short, matched tests. It is designed for teams comparing interfaces or model routes without relying on a highlight reel.

Define the shot contract #

Write the shot as a set of observable requirements before generating. A useful contract separates four layers:

1. Subject identity: clothing, face, object shape, color, and count remain recognizable.
2. Action path: the subject completes the requested motion in the correct direction and order.
3. Camera behavior: framing, movement, lens feel, and horizon remain consistent with the instruction.
4. Scene structure: walls, furniture, shadows, and background objects do not jump or melt.

Mark each requirement as mandatory or optional. A clip that looks cinematic but reverses the action direction should fail when direction is mandatory.

Use five motion probes #

Run at least one prompt from each category:

• Lateral movement: a subject crosses the frame from left to right.
• Approach or retreat: a subject moves toward or away from the camera.
• Object interaction: a hand picks up, places, opens, or closes an object.
• Camera movement: a slow pan, orbit, dolly, or push-in with a mostly static subject.
• Occlusion recovery: the subject passes behind a foreground object and reappears.

Keep duration, aspect ratio, reference image, output count, and comparable motion settings fixed. Generate at least four clips per prompt so the test measures consistency rather than one lucky result.

Score at fixed checkpoints #

Review each clip at the first frame, 25 percent, 50 percent, 75 percent, and final frame. At each checkpoint, score every layer from zero to two:

The maximum for four layers at five checkpoints is 40 points. Keep the raw layer scores instead of publishing only the total. Two clips can both score 30 while failing for different reasons.

Record failure onset #

For every rejected clip, note the earliest timestamp where the failure becomes visible. Failure onset helps distinguish routes that degrade gradually from routes that make a single abrupt transition.

Use a compact taxonomy:

• identity drift: face, clothing, color, or object design changes;
• geometry drift: body parts, props, or scene lines deform;
• trajectory error: movement follows the wrong direction or path;
• camera discontinuity: framing or camera direction jumps;
• temporal texture: flicker, crawling detail, or unstable lighting;
• interaction failure: contact between a subject and object is implausible;
• ending collapse: the final frames lose structure or freeze unnaturally.

Do not hide failed jobs, safety refusals, timeouts, or malformed exports. They belong in the denominator because they consume time and generation capacity.

Calculate usable seconds #

In addition to the continuity score, calculate:

usable-second rate = seconds meeting all mandatory requirements / total generated seconds

For example, four five-second clips produce 20 generated seconds. If only nine seconds meet every mandatory requirement, the usable-second rate is 45 percent. This exposes routes that create attractive openings but cannot sustain a complete shot.

Track review and repair time separately. A clip that needs frame trimming, speed changes, stabilization, or object cleanup may still be usable, but those corrections are part of its operational cost.

Compare routes without changing the test #

Run the same prompt set through each candidate route. Teams can include a free AI video generator workspace as one disclosed test environment while keeping prompts, output counts, checkpoints, and rejection rules identical. The link is not evidence that the workspace will score higher than another route.

Report the median continuity score, usable-second rate, rejection rate, and most common failure category. Avoid declaring a universal winner. A route that performs well on camera movement may still be weak on object interaction or occlusion recovery.

Retest after model or routing changes #

Store the date, displayed model label, settings, and source assets with every run. Repeat the protocol after a model update, routing change, prompt-enhancement change, duration change, or export-policy change. A previous score is historical evidence, not a current guarantee.

Disclosure: Written and published by PhotoArtify Team on September 14, 2026. We operate PhotoArtify and may benefit if readers use the linked video-generation workspace. This protocol is an evaluation method, not an independent endorsement or a universal performance claim.