Why explicit caching, not learned memory

The static warp+inpaint recipe that wins the static benchmarks cannot represent movers on revisit: it can only re-project a stale earlier observation, so a moving object is smeared or frozen when the camera returns. The obvious "just learn a memory" fix — a learned key/value attention memory (MemAttn) spliced into the generator — destabilises training and collapses to colored noise. What actually works is an explicit, geometry-grounded rolling 4D cache: it stores per-time point clouds and re-renders the mover from the correct time, beating the static warp on the true mover region.

+1.9 dB

mover PSNR — standard revisit

+2.7 dB

mover PSNR — drift / loop-closure

46/50

clips cache > static warp

25/25

drift clips cache wins

All deltas are on the true-mover region only (background id 0 and >35%-area static planes excluded). An earlier prototype reported a spurious +34 dB because a mask bug counted the Kubric background as a mover; the corrected mover mask gives the honest single-digit-dB win above.

1 · Static warp ghosts movers; the rolling cache renders them clean

Each panel: top-left GT · top-right static warp · bottom-left rolling 4D-cache render · bottom-right cache overlaid. The static warp smears / displaces the moving objects (it can only re-project the stale observation); the rolling cache re-renders each mover at the correct time and matches GT.

static warp vs rolling cache, mover ghosting (scene A) — Kubric revisit A — the flying cans/discs are ghost-streaked by the static warp (top-right) but crisp in the rolling-cache render (bottom-left).

static warp vs rolling cache, mover ghosting (scene B) — Kubric revisit B — the sliding bottle/shoe are frozen at the stale pose by the static warp; the cache places them at their correct return-time location.

2 · Learned KV-memory (MemAttn) is a NO-GO — it collapses to colored noise

Each strip left→right: GT · observed / early return · MemAttn attempt · rightmost = MemAttn output. Splicing a learned key/value memory into the generator destabilises the diffusion and the return frame collapses to colored noise — the same failure across scenes and away-distances, i.e. not a tuning artifact.

MemAttn collapse (scn02701 b60) — scn02701, away-bin 60°

MemAttn collapse (scn02701 b90) — scn02701, away-bin 90°

MemAttn collapse (scn02701 b120) — scn02701, away-bin 120°

MemAttn collapse (scn02708 b60) — scn02708, away-bin 60°

Takeaway. Static warp is memory-free and cannot represent movers (stale / ghost); a learned KV-memory is unstable (colored-noise collapse). An explicit geometry-grounded rolling 4D cache is what closes the loop on dynamic revisits — +1.9 dB standard / +2.7 dB drift on the true mover region, winning 46/50 clips and 25/25 drift clips. This is the scientific reason we pivoted the dynamic method to explicit caching rather than learned memory.