To check whether an AI dance follows a reference, compare the same movement phases at the same timestamps. A pleasing clip can contain the right general movement while performing it too early, too late, or in the wrong order. Write down what must match before reviewing the output; otherwise a dramatic camera move can hide a failed dance transfer.
What should I compare first?
Choose a small set of unmistakable events: the first bend, arms reaching their highest position, a turn finishing, and the return to the resting pose. Prefer events visible in both clips. An obscured foot cannot support a precise footwork claim.
Make a worksheet with columns for source time, source phase, output phase, and confidence. Mark an event as unclear when a person is hidden. Keep identity, timing, anatomy, and camera movement in separate columns rather than compressing them into one impression of quality.
If your goal is only a similar energy or mood, state that narrower requirement. If your goal is dance reproduction, matching a few attractive poses is insufficient.
What did our actual test show?
Our single-run H3 test record documents one generation with two fictional-character images and one ten-second performance reference. The output decoded successfully, but sampled comparison found these counterexamples:
| Time from the reference start | Reference group phase | Generated group phase |
|---|---|---|
| 4 seconds | Low and leaning forward | Raised and leaning backward |
| 5 seconds | Raised and leaning backward | Low and leaning forward |
| 6 seconds | Low and leaning forward | Raised with arms up |
The group performed recognizable bending and raising movements. Those observations did not establish correct choreography timing. We reviewed 80 comparison samples at eight samples per second and 20 half-second keyframes; we did not visually inspect every adjacent frame. This is one result, not a success-rate estimate or a comparison of providers.
Is the problem a fixed delay or changing drift?
Align the first shared event, then check a later event without moving the clips again. If every event has approximately the same offset, the clips may have a different start point. A trim can sometimes resolve that delivery issue. Confirm that the opening and ending still make sense after trimming.
If the offset changes or an extra movement appears between two events, the output has changed the performance sequence. Moving the whole clip cannot repair that. Speed changes may also create abrupt motion or alter the relationship between the lead and the surrounding group.
For a fast gesture, inspect neighboring frames around the event. Widely spaced screenshots can show matching end poses while missing the wrong transition. Do not claim a percentage match without a defined measurement method.
What should I change before another generation?
First isolate the intended performer in a continuous reference segment. Check whether the selected tool supports that kind of movement transfer. A general image-to-video prompt and a dedicated motion-control input have different contracts.
Then simplify one variable: shorten the motion segment, remove a competing action from the prompt, or choose a character image with compatible visible limbs. Keep the previous inputs and settings so the next result answers a specific question. Do not simultaneously change the character, camera, duration, and choreography and call the improvement reproducible.
When can I accept the clip?
Accept it when the required events remain in order, their timing meets your project's stated tolerance, and the visible transitions are usable at normal playback speed. The tolerance belongs to your project, not an invented industry standard. A loose social interpretation and a synchronized group routine need different decisions.
If the test fails, label the clip as an interpretation or reject it for the reproduction task. Preserve an unobstructed comparison worksheet with the delivery notes. A successful download proves that the generation pipeline worked; the event review determines whether the creative requirement was met.