How to read this page: original LoomData captions are unchanged. Span offsets and interval field names are normalized for consistent display; original model outputs are retained in the JSON. Twenty-five imperfect span offsets in three records were deterministically aligned to the nearest exact caption substring. RGB frames and source binary masks remain separate inputs.
Frame preview
Choose one view for the frame strips below. RGB and binary masks are shown separately.
Mask files are original LoomData selected-subject masklets paired to RGB frames at identical timestamps. A null mask link means the model could not confidently bind that text group to the supplied track. Temporal intervals are sparse clip-relative support; source-video coordinates are shown alongside them. Intervals that reach the caption clip boundary use its exact source endpoint.