Six originals generated and all366frames reviewed. No candidate passed the requested
same-performance alternate-view scope. Physically consistent multi-view motion was not obtained.
The original oblique movie remains the primary authority. No new motion SSOT or native change is adopted.
The owner requested proceeding with a source-video-conditioned front, side and rear-oblique study.
Every trial uses the same pixel-exact3s crop, originalframes24–113. The original movie, exact input
crop, provider originals, complete native-frame sheets, prompts and provenance are retained.
This tests a specific practical question: can a direct video editor change the camera while
preserving this mascot's identity, supported stance, single-hand broom task and two sweep/recovery
cycles? It does not test whether an independently generated sweeping animation looks attractive.
Two direct video editors and three strategies were used: source-video plus detailed camera prompt;
Kling with short camera instructions and an exact neutral target-view image; and Bernini with
explicit original-material/identity preservation and negative exclusions for the observed artifacts.
Kling's image guides are appearance/camera references, not motion or hidden-joint measurements.
All exact guides are displayed in the comparison page. Prior Sweep25 Wan3 video-reference trials
already failed to acquire alternate viewpoints, so those requests were not repeated.
| Candidate | Requested view | Actual result | Decision |
|---|---|---|---|
| Bernini front-r01 | Front | Frontal plush restyle; new eye highlights; two-handed grip; original two strokes not preserved | Reject |
| Kling front-r01 | Front | Original oblique camera remains; two strokes largely retained | Reject |
| Kling front-r02 | Front with image guide | Face stays oblique; visible tail reduced to a stub-like silhouette | Reject |
| Kling side-r02 | Exact side with image guide | Oblique face remains, with far eye and frontal muzzle visible | Reject |
| Bernini front-r02 | Front, material/identity constraints | Frontal flat style, but two brooms, changed face/mouth and no original two-stroke sweep | Reject |
| Bernini rear-oblique-r02 | Rear oblique | Rear/side presentation acquired; extra early broom motion and changed sweep/return sequence | Reject |
All366frames of the six completed originals were inspected in fifteen chronological sheets.
The three Bernini outputs are49frames each at16fps; the three Kling outputs are73frames
at24fps. Native frames and timestamps are preserved. This full-frame review is distinct from
successful decoding or file-integrity verification.
The three Kling originals retain the oblique camera and broadly preserve the source's two
working reaches and recoveries. Their green broom-binding feature remains identifiable.
Measured x extrema in four fixed crop-time windows are:
| Movie | Work1 | Return1 | Work2 | Return2 |
|---|---|---|---|---|
| Original source crop | 1.100s | 1.733s | 2.133s | 2.633s |
| Kling front-r01 | 1.167s | 1.750s | 2.208s | 2.750s |
| Kling front-r02 | 1.167s | 1.750s | 2.208s | 2.708s |
| Kling side-r02 | 1.208s | 1.750s | 2.208s | 2.750s |
The largest difference among these extrema is0.117s. These are binding-centroid events, not
whole-body recovery times. Flat extrema have finite plateau bands: source work1 spans0.967–1.233s
within3pixels of its minimum. The full bands are retained in timing-comparison.json. The numbers
must not be read as a precise physical joint delay, proof of identical3D motion or calibrated
synchronization. No phase shift or retiming was used to reduce differences.
The lower dark silhouette's exposed sole-bottom is constant within each studied source/Kling
clip:846/960pixels for the source and both front-labelled Kling trials,850/960for the side-labelled
trial. Rear exposed dark edge ranges are2pixels in the source,3in Kling front-r01,2in front-r02
and2in side-r02. These support an image-space stable-support observation within each clip.
They do not measure3D foot position, yaw or force. The apparent left edge moves much farther
because the broom covers it; it is not a valid stepping measurement.
The surviving oblique head/torso presentation means the Kling clips do not reveal a new pelvis,
hip or ankle view. The Bernini front clips show new surfaces only after changing the character,
tool task or performance. Their new geometry cannot resolve how the original hips or chest moved.
The upper-body work/recovery reading remains governed by the original and the Sweep27 audit.
The front-image-guided tail change cannot uniquely distinguish occlusion, reorientation and
redesign; it is an unresolved appearance change, not a measured3D shortening. The camera criterion
alone is sufficient to reject that output for this study.
Bernini front-r01 no longer has a reliably detectable green binding:39of49frames fail the50pixel
component minimum. The remaining detections are not accepted as the original feature. Front-r02
has two brooms and two possible bindings. Neither color trace is used to infer source timing,
contact or joint movement. Their raw frame boards and rejection findings are preserved.
Bernini rear-oblique-r02 changes the presentation toward the back/side and retains one broom,
one holding arm and broadly stable feet. It is the clearest new rear-oriented presentation in
this study, but its performance is different. It introduces a visible tool swing during the
first0.25s: the green binding moves up to50.70pixels in the normalized960px analysis image,
while the original projection moves0.15pixels over the same interval. These are within-view
image measurements, not equal3D lengths or a calibrated multi-view residual. The later brush
direction changes and returns also do not reproduce the source two-stroke sequence.
All49frames were reviewed; native-sizeframes2and20 and seven feature-mask keys were inspected.
The brush crosses the projected tail-side silhouette around1.25s. Its apparent height and
overlap cannot alone prove penetration, lifting off the floor or a particular3D sweep direction.
This candidate may help imagine a rear appearance, but it cannot supply accepted ankle, hip
or torso motion for the original. It is rejected for the same-performance scope.
The completed camera edits demonstrate a practical failure mode: keeping the performance tends
to retain the source viewpoint, while acquiring a frontal view can invent a different performance.
This is a finding about these submitted cases, not a claim that every video model must fail.
No rejected output should be relabelled as a physically consistent multi-view SSOT.
Foot yaw, hip accommodation and the pelvis/chest angle split remain unresolved. Continue to use
Sweep27's controlled seven-pose common-frame versus articulated-rig experiment if native work
resumes. The generated candidates introduce no new authoritative joint targets. See
Rig implications and Methods.
The submitted batch comprises three Bernini requests estimated at$0.245each and three Kling
requests estimated at$0.504each, total$2.247. Each estimate was shown before submission through
mediagen. The provider's actual billed amount is not returned; the generic estimate is explicitly
marked caller-supplied and unverified. Four unused round1 side/rear specifications were prepared
but never submitted. No duplicate paid request was made while a job was pending.
The full original, Sweep25 native and adopted releases are preserved. No mechanical fitting,
new native controls, production adoption, independent model review or subagent work occurs here.