MCSDF — knownness videos (forward ctx18)
Memory built from the first 18 frames only · geometry from VGGT · 81 frames @ 12fps
Each video: GT (camera view) | z_mem (memory reprojection) | knownness k (viridis, aligned overlay:
yellow = seen / blue = unseen). Watch the bar: green 0–17 = context, orange 18–80 = generated.
As the camera moves past the 18 context frames, z_mem thins out and k drops in the newly-revealed regions — that low-k is exactly where the diffusion model generates from noise.