luuuulinnnn's picture
Interim viz build: multi-visit return strip now GROUPS by pose. Each pose renders as a labeled box (Pose A = first observed frame, Pose B = mid observed frame) that LEADS with the Observed GT as the reference target on the left, followed by that method's Visit 1 / Visit 2 / Visit 3 of the SAME pose, each captioned 'Visit k'. Added the caption 'Each Observed GT is the target its same-pose visits should reproduce (GT-PSNR = visit vs this GT; self-PSNR = visit vs Visit 1)', so the obs<->visits correspondence is now unmistakable. Preserves the 倍速 speed control, lockstep sync, native-res grids, the moving-back single-pose strips, the 'Why explicit caching' motivation section, and all prior sections. Deferred (render gracefully, not yet landed): dynamic-region mover-only PSNR (n/a) and the DiT+VACE-full ablation arm (pending). [root]
4774156 verified
Raw
History Blame Contribute Delete
6.4 kB
<!-- BEGIN mechanism (motivation: explicit cache vs learned memory) -->
<section class="block">
<section id="mechanism">
<style>
#mechanism{font-family:-apple-system,BlinkMacSystemFont,"Segoe UI",Roboto,Helvetica,Arial,sans-serif;color:#1b1f24;max-width:100%;margin:0 auto;padding:8px 4px 40px;}
#mechanism h2{font-size:1.55rem;margin:.2em 0 .15em;font-weight:750;letter-spacing:-.01em;}
#mechanism h3{font-size:1.12rem;margin:1.4em 0 .3em;font-weight:650;color:#0b2545;border-bottom:2px solid #e3e8ef;padding-bottom:3px;}
#mechanism .lede{font-size:.96rem;line-height:1.6;color:#3a4452;max-width:70rem;margin:.3em 0 .6em;}
#mechanism .lede b{color:#0b2545;}
#mechanism .win{color:#0a7d2c;font-weight:750;}
#mechanism .nogo{color:#9a2b15;font-weight:750;}
#mechanism .kpis{display:flex;flex-wrap:wrap;gap:12px;margin:.6em 0 .2em;}
#mechanism .kpi{border:1px solid #dfe6ee;border-radius:10px;background:#fff;padding:8px 14px;min-width:150px;box-shadow:0 1px 2px rgba(20,30,55,.04);}
#mechanism .kpi .v{font-size:1.28rem;font-weight:800;color:#0a7d2c;letter-spacing:-.01em;}
#mechanism .kpi .k{font-size:.72rem;font-weight:700;letter-spacing:.05em;text-transform:uppercase;color:#8a929c;margin-top:1px;}
#mechanism figure{margin:.5em 0 .3em;}
#mechanism figure img{width:100%;height:auto;border:1px solid #e3e8ef;border-radius:8px;display:block;}
#mechanism .two{display:grid;grid-template-columns:1fr 1fr;gap:14px;}
#mechanism .four{display:grid;grid-template-columns:1fr 1fr;gap:12px;}
#mechanism figcaption{font-size:.8rem;color:#6b7480;line-height:1.5;margin-top:4px;}
#mechanism .panelkey{font-size:.8rem;color:#3a4452;background:#f6f8fb;border-left:3px solid #cdd6e0;padding:6px 10px;border-radius:0 6px 6px 0;margin:.3em 0 .6em;}
#mechanism .concl{font-size:.95rem;line-height:1.6;color:#1b1f24;background:#f2f8f3;border:1px solid #cde8d3;border-radius:10px;padding:11px 15px;margin:1.1em 0 .2em;max-width:70rem;}
@media(max-width:720px){#mechanism .two,#mechanism .four{grid-template-columns:1fr;}}
</style>
<h2>Why explicit caching, not learned memory</h2>
<p class="lede">The static warp+inpaint recipe that wins the <b>static</b> benchmarks cannot represent
<b>movers</b> 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 &mdash; a learned
key/value attention memory (MemAttn) spliced into the generator &mdash; <span class="nogo">destabilises
training and collapses to colored noise</span>. What actually works is an <b>explicit, geometry-grounded
rolling 4D&nbsp;cache</b>: it stores per-time point clouds and re-renders the mover from the correct
time, beating the static warp on the true mover region.</p>
<div class="kpis">
<div class="kpi"><div class="v">+1.9&nbsp;dB</div><div class="k">mover PSNR &mdash; standard revisit</div></div>
<div class="kpi"><div class="v">+2.7&nbsp;dB</div><div class="k">mover PSNR &mdash; drift / loop-closure</div></div>
<div class="kpi"><div class="v">46/50</div><div class="k">clips cache &gt; static warp</div></div>
<div class="kpi"><div class="v">25/25</div><div class="k">drift clips cache wins</div></div>
</div>
<p class="lede" style="font-size:.82rem;color:#6b7480;">All deltas are on the <b>true-mover region only</b>
(background id&nbsp;0 and &gt;35%-area static planes excluded). An earlier prototype reported a spurious
+34&nbsp;dB because a mask bug counted the Kubric background as a mover; the corrected mover mask gives
the honest single-digit-dB win above.</p>
<h3>1 &middot; Static warp ghosts movers; the rolling cache renders them clean</h3>
<div class="panelkey">Each panel: <b>top-left</b> GT &middot; <b>top-right</b> static warp &middot;
<b>bottom-left</b> rolling 4D-cache render &middot; <b>bottom-right</b> 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.</div>
<div class="two">
<figure><img src="assets/mechanism/warp_vs_cache_1.png" alt="static warp vs rolling cache, mover ghosting (scene A)">
<figcaption>Kubric revisit A &mdash; the flying cans/discs are ghost-streaked by the static warp
(top-right) but crisp in the rolling-cache render (bottom-left).</figcaption></figure>
<figure><img src="assets/mechanism/warp_vs_cache_2.png" alt="static warp vs rolling cache, mover ghosting (scene B)">
<figcaption>Kubric revisit B &mdash; the sliding bottle/shoe are frozen at the stale pose by the
static warp; the cache places them at their correct return-time location.</figcaption></figure>
</div>
<h3>2 &middot; Learned KV-memory (MemAttn) is a NO-GO &mdash; it collapses to colored noise</h3>
<div class="panelkey">Each strip left&rarr;right: GT &middot; observed / early return &middot; MemAttn
attempt &middot; <b>rightmost = MemAttn output</b>. Splicing a learned key/value memory into the generator
destabilises the diffusion and the return frame collapses to <span class="nogo">colored noise</span> &mdash;
the same failure across scenes and away-distances, i.e. not a tuning artifact.</div>
<div class="four">
<figure><img src="assets/mechanism/memattn_collapse_1.png" alt="MemAttn collapse (scn02701 b60)">
<figcaption>scn02701, away-bin 60&deg;</figcaption></figure>
<figure><img src="assets/mechanism/memattn_collapse_2.png" alt="MemAttn collapse (scn02701 b90)">
<figcaption>scn02701, away-bin 90&deg;</figcaption></figure>
<figure><img src="assets/mechanism/memattn_collapse_3.png" alt="MemAttn collapse (scn02701 b120)">
<figcaption>scn02701, away-bin 120&deg;</figcaption></figure>
<figure><img src="assets/mechanism/memattn_collapse_4.png" alt="MemAttn collapse (scn02708 b60)">
<figcaption>scn02708, away-bin 60&deg;</figcaption></figure>
</div>
<div class="concl"><b>Takeaway.</b> Static warp is memory-free and cannot represent movers (stale /
ghost); a learned KV-memory is unstable (colored-noise collapse). An <b>explicit geometry-grounded
rolling 4D cache</b> is what closes the loop on dynamic revisits &mdash; <span class="win">+1.9&nbsp;dB
standard / +2.7&nbsp;dB drift on the true mover region, winning 46/50 clips and 25/25 drift clips</span>.
This is the scientific reason we pivoted the dynamic method to explicit caching rather than learned
memory.</div>
</section>
</section>
<!-- END mechanism -->