Per-token noise level = knownness · memory built from the first 18 frames only · VGGT-consistent geometry
What this shows. The memory is built from only the first 18 context frames, so the far/peripheral
regions are genuinely unseen and must be generated — real scene extrapolation, not reconstruction. The
knownness map k (which drives each token's noise level) now varies meaningfully: std 0.02 → 0.18,
yellow where the context saw it, blue where it didn't. Geometry is from VGGT (cross-frame-consistent),
which de-shatters the memory the model conditions on. Two orthogonal fixes underneath: VGGT geometry (clean z_mem)
+ the per-token Δt / x0-projection sampler (sharpness).
00534f58 pier / covered walkway
Camera trajectory — blue = the 18 context frames the memory is built from (SEEN); red = target frames we generate (UNSEEN); ★ = frame 40, the far/turn point.Out-and-back: frame 0 (context) and frame 80 (return) are the same viewpoint → identical content (a memory/revisit test); frame 40 is a completely different far view.
Knownness map k as a heatmap across the video. Blue = unseen (must be generated), yellow = seen (covered by the context memory). Note it VARIES — sky/edges blue, corridor floor yellow.
A KNOWN region (high k): z_mem carries the real content → close to GT (near-identical).An UNKNOWN region (low k): z_mem is empty → the model must generate it (completely different from an empty memory).
0003dc82 plaza / blue seating
Camera trajectory — blue = the 18 context frames the memory is built from (SEEN); red = target frames we generate (UNSEEN); ★ = frame 40, the far/turn point.Out-and-back: frame 0 (context) and frame 80 (return) are the same viewpoint → identical content (a memory/revisit test); frame 40 is a completely different far view.
Knownness map k as a heatmap across the video. Blue = unseen (must be generated), yellow = seen (covered by the context memory). Note it VARIES — sky/edges blue, corridor floor yellow.
A KNOWN region (high k): z_mem carries the real content → close to GT (near-identical).An UNKNOWN region (low k): z_mem is empty → the model must generate it (completely different from an empty memory).
