Owner requested proceeding with the agreed multi-angle generation and analysis strategy on2026-10-04.
The original oblique sweep remains the primary motion authority. New videos are supplemental hypotheses, not physical ground truth.
No mechanical fitting or native rig change is included in this stage.
Original: output/ssot_candidates/sweep23_fixedshoulder_20261003/r01/original/output.mp4,
SHA256 21f8a76609d8bbaa3a100592d7f38ece48f469d622f0130e4532f635bd318fd4.
The input uses sourceframes24–113 inclusive,0.8–3.8s,90frames at30fps,960x960.
A lossless H264 temporal crop retains exact decoded pixels for all90frames; no retiming/interpolation/spatial editing.
CropSHA256 43478a044fe90c08f3ac2e904d25fb6397e8d6417a5841b801cbbfbdf17ac6f4.
Lineage and all90 pixel comparisons are in output/ssot_candidates/sweep28_multiview_20261004/source/lineage.json.
The clip contains two reaches and two recoveries; the third stroke and final settle are outside this pilot.
Sweep25 already tried Wan3 with the liked reference video: front-r01,side-r01 andfront-r02 retained the oblique view.
Front-r02 included a correct front image but still failed camera coverage. Selected Sweep25 views instead used
image endpoints and semantic action prompts; they are not source-video-conditioned synchronized views.
Read ../sweep25/SOURCE_REVIEW.md. This study tests direct video-editing endpoints and checks actual event timing;
it does not repeat the failed Wan3 video-reference approach or relabel independent motion as synchronized.
Round1 compares direct video editing: Bernini-R front and Kling O3 Pro front. Both receive the same entire crop.
Bernini:49frames requested at16fps,848px long edge; anticipated3.0625s x$0.08/s = $0.245.
Kling: source3s x$0.168/s = $0.504. Initial total$0.749.
The advertised rates were checked on2026-10-04 at:
- https://fal.ai/models/fal-ai/bernini-r/edit-video
- https://fal.ai/models/fal-ai/kling-video/o3/pro/video-to-video/edit
Both estimates were submitted to mediagen before paid generation. Generic endpoint estimates are caller-supplied,
not internally price-verified. Current spend_control=platform; no owner approval token is required.
fal authenticated probe: status404,ok=true (nil-UUID status test). No credentials are copied or accessed by the caller.
Select the more faithful method for exact side and rear-oblique views. If neither preserves timing/identity/support,
record failures and adjust the method/prompt using those specific observations. Do not treat an attractive new performance
as another camera of the original. View instructions are targets, never proof of calibrated angle.
Bind every decision to the untouched provider movie hash and provenance. Decode all original timestamps.
Inspect every frame for viewpoint, anatomical identity, fixed support, body proportions, hand/shaft contact and visibility.
Compare work/recovery timing to source1.1/1.7/2.2/2.7s in crop time; preserve actual playback rate for this check.
Screen coordinates across different cameras cannot be compared as equal3D joints. Measure within-view image features,
then compare event timing and semantic relationships. Distinguish camera drift, body rotation and foot displacement.
A passing supplemental result must have no unresolved blocking issue in its stated scope. Keep rejected results for comparison.
Do not infer a pelvis/chest or paw angle from a moving covering boundary, nor treat matching generated views as independent evidence.
Provide a source/candidate comparison player, complete frame sheets, measurements and explicit decisions/limitations.
Continue the existing owner-authorized public review delivery. Preserve Sweep25 native and all adopted releases.
Six requests completed; all366frames reviewed. All six are rejected for the original same-performance alternate-view scope.
Round2 added exact front/side image guides to Kling; round3 added material/identity constraints to Bernini front/rear requests.
Rear presentation changed, but the source timing/action did not survive. No further generation is pending and no generated view
is registered as motion SSOT. Total caller-estimated cost$2.247, actual billing unavailable. Read REPORT.md and METHODS.md.