• A 12-Shot Motion Continuity Protocol for AI Video Generators
  • Test Setup
  • Five Continuity Scores
  • Weight the Failures by Production Cost
  • Use Failure Tags, Not Only Scores
  • Retest the Weakest Constraint
  • Reporting Template

§A 12-Shot Motion Continuity Protocol for AI Video Generators

Short AI video demos often hide the failure that matters most in production: whether motion remains physically and visually continuous from the first frame to the last. A clip can look impressive as a thumbnail and still fail when a subject changes scale, a prop disappears, or the camera movement stops matching the scene.

This protocol turns continuity review into a repeatable 12-shot test. It is designed for image-to-video and text-to-video systems, and it can be run without a private benchmark dataset.

§Test Setup

Use one subject reference, one aspect ratio, one clip duration, and one output resolution throughout the test. Disable post-processing that changes frame interpolation or camera motion after generation. Record the model version and every non-default parameter.

Create four scene families and generate three clips in each family:

The first nine shots isolate one source of motion. The final three combine sources and reveal whether the system can preserve several constraints at once.

§Five Continuity Scores

Score every clip from 0 to 4 on each dimension. A score of 4 means no visible failure at normal playback speed; 0 means the clip cannot be used.

1. Identity continuity: facial structure, clothing, product shape, colors, and defining marks remain stable.
2. Geometry continuity: limbs, object edges, repeated patterns, and background structures deform plausibly.
3. Motion path continuity: velocity and direction change smoothly instead of jumping between unrelated trajectories.
4. Camera continuity: parallax, perspective, framing, and camera acceleration agree with the requested move.
5. Temporal cleanup: the final frames do not collapse, smear, duplicate subjects, or introduce an abrupt style shift.

The maximum raw score is 240 points: 12 clips multiplied by five dimensions and four points. Convert it to a percentage for comparisons across test runs.

§Weight the Failures by Production Cost

Not every defect has the same consequence. Apply a multiplier after the raw review:

• Multiply identity and geometry failures by 1.5 for product, character, or spokesperson work.
• Multiply camera failures by 1.5 when the clip must match an edit or storyboard.
• Multiply final-frame failures by 1.25 when clips will be extended or chained.
• Leave minor background texture changes at their raw value unless the environment carries brand information.

This prevents a large number of harmless texture variations from hiding one expensive identity failure.

§Use Failure Tags, Not Only Scores

Attach one or more short tags to each failed clip. A useful starter taxonomy is identity-drift, object-loss, limb-rewrite, background-morph, motion-jump, camera-stall, parallax-error, occlusion-reset, and end-frame-collapse.

Tags reveal patterns that an average score cannot. For example, two models may both score 78%, while one fails mainly on low-cost background details and the other repeatedly changes the subject’s face.

§Retest the Weakest Constraint

After the first 12 clips, change only one prompt instruction or generation control. If camera stalls dominate, make the camera path observable: specify start distance, end distance, direction, and subject framing. If identity drift dominates, reduce simultaneous motion and describe the invariant subject features before the action.

Run the same 12 shots again. A useful change should improve the targeted failure rate without creating a new dominant defect.

§Reporting Template

For every test run, publish:

• model and version;
• generation date;
• input image provenance or text prompt;
• resolution, duration, aspect ratio, seed policy, and motion controls;
• per-clip scores and failure tags;
• rejected clips, not only selected examples;
• total generation count and approximate cost per accepted clip.

The cost per accepted clip is often more operationally useful than the price of one generation. It captures the reruns, review time, and manual fixes needed to reach a usable result.

Disclosure: this protocol is maintained by the PhotoArtify Team. We operate PhotoArtify and may benefit if readers visit or use it. The protocol is a reproducible evaluation method, not an independent endorsement or a claim that one model will fit every workflow.