Baylands, take T4 (cycling, 4K100), 20 seconds from the drones' own pixels

How it's made. A body model fitted on every frame to all three drones (2D keypoints at each drone's own exposure time, silhouettes, triangulated joints) gives the shape. Each pixel of it is coloured from the drones' 4K frames at that instant, blended by how close each drone's view is to yours, how far inside the athlete's silhouette the pixel falls, and whether that drone can actually see the point. Nothing is generated. The body is then refined so its outline matches the athlete in every drone view (silhouette overlap 0.75–0.80 before, 0.94–0.97 after). The field is a ground map and a tree-line panorama baked from 270 drone frames with the athlete and their shadow removed; the shadow you see is cast along the real sun (azimuth 250°, 20° up, from the capture time and GPS).

Generated fill. Angles none of the three drones saw at that moment (often the front, when all three follow from behind) are filled from six views generated by 4DAnyone, a research video model, conditioned on the measured three-drone body and the drones' real footage, and placed as virtual cameras that move with the rider. The drones' own pixels win wherever a drone saw the point. Turn it off with Generated fill to see the drones alone. Research benchmark: the model's weights are not licensed for commercial use.

Where it breaks. The free camera is limited to the heights the drones covered; hands and other small, fast parts are the least accurate. Drone 3 has no pose for the first 4.6 s, so early frames use two drones. The field is flat by construction, so anything standing on it apart from the athlete is flattened.