sections/memory.html

<section class="block">
<section id="memory">
<style>
#memory{font-family:-apple-system,BlinkMacSystemFont,"Segoe UI",Roboto,Helvetica,Arial,sans-serif;color:#1b1f24;max-width:100%;margin:0 auto;padding:8px 4px 40px;}
#memory h2{font-size:1.55rem;margin:.2em 0 .15em;font-weight:700;letter-spacing:-.01em;}
#memory h3{font-size:1.12rem;margin:1.4em 0 .25em;font-weight:650;color:#0b2545;border-bottom:2px solid #e3e8ef;padding-bottom:3px;}
#memory .sub{font-size:.86rem;color:#5b6470;margin:.1em 0 .7em;line-height:1.4;}
#memory .tblwrap{overflow-x:auto;border:1px solid #e3e8ef;border-radius:7px;margin:.3em 0 .4em;}
#memory table{border-collapse:collapse;width:100%;font-size:.8rem;font-variant-numeric:tabular-nums;}
#memory thead th{position:sticky;top:0;background:#0b2545;color:#fff;font-weight:600;padding:6px 8px;text-align:right;white-space:nowrap;z-index:2;border-bottom:2px solid #08203d;font-size:.76rem;}
#memory thead th.lt{text-align:left;}
#memory thead th.grp{background:#123a63;text-align:center;border-left:1px solid #2a4d72;border-right:1px solid #2a4d72;font-size:.72rem;letter-spacing:.02em;text-transform:uppercase;}
#memory thead th.star{background:#5a3a73;}
#memory thead th.bin{border-left:1px solid #2a4d72;}
#memory tbody td{padding:5px 8px;text-align:right;white-space:nowrap;border-bottom:1px solid #eef1f5;}
#memory tbody td.lt{text-align:left;}
#memory tbody tr:nth-child(even){background:#fafbfc;}
#memory tbody tr.ours{background:#fff6e0;}
#memory tbody tr.ours:nth-child(even){background:#fdf0d2;}
#memory tbody tr.oursft{background:#f3f7ff;}
#memory tbody tr.oursft:nth-child(even){background:#ecf2fc;}
#memory tbody tr:hover{background:#eef4fb;}
#memory tbody td strong{color:#0a7d2c;font-weight:750;}
#memory .na{color:#b6bdc6;}
#memory .run{color:#b06f00;font-style:italic;font-size:.74rem;}
#memory .nbadge{color:#7a828c;font-weight:400;font-size:.72rem;}
#memory .mname{font-size:.78rem;}
#memory .tag{display:inline-block;border-radius:4px;padding:1px 6px;font-size:.66rem;font-weight:700;letter-spacing:.02em;vertical-align:middle;margin-left:6px;}
#memory .tag.ours{background:#f0a500;color:#231a00;}
#memory .tag.pri{background:#0a7d2c;color:#fff;}
#memory .tag.std{background:#15539e;color:#fff;}
#memory .ftlabel{display:inline-block;background:#e6efff;color:#15539e;border-radius:4px;padding:0 5px;font-size:.7rem;font-weight:600;margin-left:4px;vertical-align:middle;}
#memory .psbadge{display:inline-block;background:#eef3f7;color:#5a6b78;border-radius:4px;padding:0 5px;font-size:.66rem;font-weight:600;vertical-align:middle;margin-left:4px;}
#memory .note{font-size:.78rem;line-height:1.45;color:#54606e;background:#f6f8fb;border-left:3px solid #c2cfde;padding:7px 11px;border-radius:0 5px 5px 0;margin:.2em 0 1.1em;}
#memory .note.warn{background:#fff7ed;border-left-color:#e0922f;color:#6b4a16;}
#memory .note b{color:#1b1f24;}
#memory .head{background:#f3eef9;border-left:3px solid #7a4f9a;padding:9px 13px;border-radius:0 6px 6px 0;font-size:.86rem;color:#3a2b4a;margin:.2em 0 1em;}
#memory .head b{color:#2a1a3a;}
</style>

<h2>Memory / revisit &mdash; a 4th axis (NEW)</h2>
<div class="head"><b>Can a method REMEMBER what it has already seen?</b> Single-shot &amp; sliding-window
generators have no persistent world state &mdash; when the camera <b>revisits</b> an observed pose or
<b>returns</b> after wandering into unseen space, they re-hallucinate it rather than recall it. These two
new benchmark families isolate exactly that. Both score <b>only on real GT</b>: the revisit/return legs
carry true ground-truth frames; the synthetic away-legs are masked out upstream.</div>
<div class="sub"><b>Two families &times; three sources</b> (Kubric-4D clean-GT primary, DL3DV static,
SpatialVID dynamic). <b>M1 multi_visit</b> = REVISIT-CONSISTENCY (GT-fidelity at a revisited observed
pose + cross-visit self-consistency drift). <b>M2 moving_back</b> = RETURN-FIDELITY (quality back at the
observed pose after a round-trip, and the PSNR <i>degradation</i> vs the pre-departure observed baseline
&mdash; the forgetting signal). <span class="run">pending</span> = method has outputs but is not yet
scored on this family; methods with no run are omitted. <b>9 methods scored per benchmark now</b>
(incl. the dedicated spatial-memory model <b>spmem</b> and the warp-conditioned Wan2.1-14B variant
<b>FrameCrafter</b>, both on <b>all six</b> memory benchmarks &mdash; both families
&times; all three scene types); FlexWorld/TrajCrafter/ViewExtrapolator/Voyager and spmem on the still-generating
long-horizon set remain <span class="run">pending</span>.</div>

<div class="note warn">&#9888; <b>Headline &mdash; explicit geometric caching, not learned memory, enables return-fidelity.
Now confirmed across all three scene types.</b> The <b>moving-back degradation</b> column (return-PSNR drop vs the
method&rsquo;s own pre-departure observed baseline; smaller = better memory) is the story, and the same split now
holds across <b>static-real (DL3DV), synthetic (Kubric-4D clean-GT), AND dynamic-real (SpatialVID)</b> &mdash; so
it is a real capability gap, not a domain or render artifact.
<ul style="margin:.5em 0 .2em;padding-left:1.3em;line-height:1.5;">
<li><b>Geometric memory WORKS.</b> Per-scene 4D-GS (Shape-of-Motion) <i>stores the whole scene</i> and barely
degrades &mdash; <b>&minus;0.8&nbsp;dB (DL3DV) / &minus;1.0&nbsp;dB (SpatialVID)</b>. Among <i>forward
generators</i>, <b>GEN3C</b> &mdash; which keeps an explicit 3D point cache &mdash; is the best by a wide margin
and is the only generator that stays low on <b>every</b> scene type: <b>&minus;17.0&nbsp;dB (DL3DV) /
&minus;13.8&nbsp;dB (Kubric) / &minus;17.5&nbsp;dB (SpatialVID)</b> (return-PSNR 17.3 / 22.4 / 17.3).</li>
<li><b>Learned / neural spatial memory does NOT (yet).</b> <b>spmem</b> &mdash; the dedicated spatial-memory paper,
now scored on all six memory benchmarks &mdash; collapses <b>&minus;30.0&nbsp;dB (DL3DV) / &minus;24.1&nbsp;dB
(Kubric) / &minus;32.0&nbsp;dB (SpatialVID)</b>, among the worst on all three. The collapse is <i>uniform across
static, synthetic and dynamic</i>, so it is <b>not a domain gap</b>; the learned memory simply isn&rsquo;t
recalling the observed pose &mdash; it forgets exactly like the memoryless single-shot generators.</li>
<li><b>The memoryless single-shot generators all collapse in the same band.</b> Lyra&nbsp;2.0 (&minus;28.1 / &minus;19.4
/ &minus;27.6), Ours(base&nbsp;VACE) (&minus;28.2 / &minus;21.7 / &minus;29.2), NVS-Solver (&minus;31.9 / &minus;21.0
/ &minus;31.0), ViewCrafter (&minus;28.5 / &minus;25.4 / &minus;32.5)&nbsp;dB &mdash; the <b>~&minus;20&nbsp;to
&minus;32&nbsp;dB</b> band that motivates this axis: today&rsquo;s strongest extrapolators have <b>no working
memory</b>, and spmem sits squarely inside it.</li>
</ul>
<b>Honest caveat:</b> Matrix-Game&nbsp;3.0&rsquo;s tiny degrade (&minus;1.8 DL3DV / &minus;9.5 Kubric / &minus;3.7
SpatialVID&nbsp;dB) is <b>not memory</b> &mdash; it is an artifact of <i>uniformly-low quality</i> (return-PSNR
&asymp;&nbsp;11 with observed-baseline &asymp;&nbsp;13&ndash;21 on all three), so there is little to forget. Always
read the degrade column together with the absolute return-PSNR.</div>

<h3>Headline result &mdash; return-degradation across all three scene types <span class="tag pri">CAMPAIGN HEADLINE</span></h3>
<div class="sub">The single number that carries this axis: <b>return-PSNR degradation</b> (drop vs the method&rsquo;s own
pre-departure observed baseline &mdash; <b>lower bars = better memory</b>). Six key methods &times; three scene types.
The shape is identical everywhere: <b>geometric caching stays near zero, every generator (learned-memory spmem
included) collapses</b>. All values are real-GT means read directly from the moving-back JSONs.</div>
<div class="tblwrap" style="border:none;background:#fff"><svg viewBox="0 0 760 300" width="100%" preserveAspectRatio="xMidYMid meet" role="img" aria-label="Return-PSNR degradation (dB) by method across three scene types — lower is better memory" style="font-family:-apple-system,Segoe UI,Roboto,Helvetica,Arial,sans-serif;background:#fff"><text x="46" y="18" font-size="13" font-weight="700" fill="#0b2545">Return-PSNR degradation after a round-trip (dB)</text><text x="748" y="18" font-size="9.5" text-anchor="end" fill="#8a929c">&#8595; lower = better memory</text><line x1="46" y1="242.0" x2="748" y2="242.0" stroke="#eef1f5" stroke-width="1"/><text x="42" y="245.0" font-size="9" text-anchor="end" fill="#6b7480">0</text><line x1="46" y1="213.1" x2="748" y2="213.1" stroke="#eef1f5" stroke-width="1"/><text x="42" y="216.1" font-size="9" text-anchor="end" fill="#6b7480">5</text><line x1="46" y1="184.3" x2="748" y2="184.3" stroke="#eef1f5" stroke-width="1"/><text x="42" y="187.3" font-size="9" text-anchor="end" fill="#6b7480">10</text><line x1="46" y1="155.4" x2="748" y2="155.4" stroke="#eef1f5" stroke-width="1"/><text x="42" y="158.4" font-size="9" text-anchor="end" fill="#6b7480">15</text><line x1="46" y1="126.6" x2="748" y2="126.6" stroke="#eef1f5" stroke-width="1"/><text x="42" y="129.6" font-size="9" text-anchor="end" fill="#6b7480">20</text><line x1="46" y1="97.7" x2="748" y2="97.7" stroke="#eef1f5" stroke-width="1"/><text x="42" y="100.7" font-size="9" text-anchor="end" fill="#6b7480">25</text><line x1="46" y1="68.9" x2="748" y2="68.9" stroke="#eef1f5" stroke-width="1"/><text x="42" y="71.9" font-size="9" text-anchor="end" fill="#6b7480">30</text><line x1="46" y1="40.0" x2="748" y2="40.0" stroke="#eef1f5" stroke-width="1"/><text x="42" y="43.0" font-size="9" text-anchor="end" fill="#6b7480">35</text><line x1="46" y1="242.0" x2="748" y2="242.0" stroke="#c2ccd6" stroke-width="1.2"/><text x="13" y="141.0" font-size="9.5" text-anchor="middle" fill="#54606e" transform="rotate(-90 13 141.0)">degrade (dB)</text><rect x="60.0" y="237.2" width="25.2" height="4.8" fill="#1f77b4" opacity="0.9"><title>Shape-of-Motion · DL3DV (static-real): -0.8 dB</title></rect><text x="72.6" y="235.2" font-size="7.5" text-anchor="middle" fill="#3a4452">1</text><rect x="89.3" y="239.0" width="25.2" height="3" fill="#d9dee5"/><text x="101.9" y="236.0" font-size="7" text-anchor="middle" fill="#9aa3ad">n/a</text><rect x="118.5" y="236.2" width="25.2" height="5.8" fill="#2ca02c" opacity="0.9"><title>Shape-of-Motion · SpatialVID (dyn-real): -1.0 dB</title></rect><text x="131.1" y="234.2" font-size="7.5" text-anchor="middle" fill="#3a4452">1</text><text x="104.5" y="258.0" font-size="9.5" text-anchor="middle" fill="#3a4452" font-weight="600">Shape-of-Motion</text><rect x="177.0" y="143.7" width="25.2" height="98.3" fill="#1f77b4" opacity="0.9"><title>GEN3C · DL3DV (static-real): -17.0 dB</title></rect><text x="189.6" y="141.7" font-size="7.5" text-anchor="middle" fill="#3a4452">17</text><rect x="206.3" y="162.1" width="25.2" height="79.9" fill="#ff7f0e" opacity="0.9"><title>GEN3C · Kubric (synthetic): -13.8 dB</title></rect><text x="218.9" y="160.1" font-size="7.5" text-anchor="middle" fill="#3a4452">14</text><rect x="235.5" y="140.9" width="25.2" height="101.1" fill="#2ca02c" opacity="0.9"><title>GEN3C · SpatialVID (dyn-real): -17.5 dB</title></rect><text x="248.1" y="138.9" font-size="7.5" text-anchor="middle" fill="#3a4452">18</text><text x="221.5" y="258.0" font-size="9.5" text-anchor="middle" fill="#3a4452" font-weight="600">GEN3C</text><rect x="294.0" y="79.4" width="25.2" height="162.6" fill="#1f77b4" opacity="0.9"><title>Ours · VACE · DL3DV (static-real): -28.2 dB</title></rect><text x="306.6" y="77.4" font-size="7.5" text-anchor="middle" fill="#3a4452">28</text><rect x="323.3" y="117.0" width="25.2" height="125.0" fill="#ff7f0e" opacity="0.9"><title>Ours · VACE · Kubric (synthetic): -21.6 dB</title></rect><text x="335.9" y="115.0" font-size="7.5" text-anchor="middle" fill="#3a4452">22</text><rect x="352.5" y="73.4" width="25.2" height="168.6" fill="#2ca02c" opacity="0.9"><title>Ours · VACE · SpatialVID (dyn-real): -29.2 dB</title></rect><text x="365.1" y="71.4" font-size="7.5" text-anchor="middle" fill="#3a4452">29</text><text x="338.5" y="258.0" font-size="9.5" text-anchor="middle" fill="#3a4452" font-weight="600">Ours · VACE</text><rect x="411.0" y="79.7" width="25.2" height="162.3" fill="#1f77b4" opacity="0.9"><title>Lyra 2.0 · DL3DV (static-real): -28.1 dB</title></rect><text x="423.6" y="77.7" font-size="7.5" text-anchor="middle" fill="#3a4452">28</text><rect x="440.3" y="130.2" width="25.2" height="111.8" fill="#ff7f0e" opacity="0.9"><title>Lyra 2.0 · Kubric (synthetic): -19.4 dB</title></rect><text x="452.9" y="128.2" font-size="7.5" text-anchor="middle" fill="#3a4452">19</text><rect x="469.5" y="82.6" width="25.2" height="159.4" fill="#2ca02c" opacity="0.9"><title>Lyra 2.0 · SpatialVID (dyn-real): -27.6 dB</title></rect><text x="482.1" y="80.6" font-size="7.5" text-anchor="middle" fill="#3a4452">28</text><text x="455.5" y="258.0" font-size="9.5" text-anchor="middle" fill="#3a4452" font-weight="600">Lyra 2.0</text><rect x="528.0" y="68.9" width="25.2" height="173.1" fill="#1f77b4" opacity="0.9"><title>spmem · DL3DV (static-real): -30.0 dB</title></rect><text x="540.6" y="66.9" font-size="7.5" text-anchor="middle" fill="#3a4452">30</text><rect x="557.3" y="103.0" width="25.2" height="139.0" fill="#ff7f0e" opacity="0.9"><title>spmem · Kubric (synthetic): -24.1 dB</title></rect><text x="569.9" y="101.0" font-size="7.5" text-anchor="middle" fill="#3a4452">24</text><rect x="586.5" y="57.3" width="25.2" height="184.7" fill="#2ca02c" opacity="0.9"><title>spmem · SpatialVID (dyn-real): -32.0 dB</title></rect><text x="599.1" y="55.3" font-size="7.5" text-anchor="middle" fill="#3a4452">32</text><text x="572.5" y="258.0" font-size="9.5" text-anchor="middle" fill="#3a4452" font-weight="600">spmem</text><rect x="645.0" y="57.7" width="25.2" height="184.3" fill="#1f77b4" opacity="0.9"><title>NVS-Solver · DL3DV (static-real): -31.9 dB</title></rect><text x="657.6" y="55.7" font-size="7.5" text-anchor="middle" fill="#3a4452">32</text><rect x="674.3" y="120.6" width="25.2" height="121.4" fill="#ff7f0e" opacity="0.9"><title>NVS-Solver · Kubric (synthetic): -21.0 dB</title></rect><text x="686.9" y="118.6" font-size="7.5" text-anchor="middle" fill="#3a4452">21</text><rect x="703.5" y="63.3" width="25.2" height="178.7" fill="#2ca02c" opacity="0.9"><title>NVS-Solver · SpatialVID (dyn-real): -31.0 dB</title></rect><text x="716.1" y="61.3" font-size="7.5" text-anchor="middle" fill="#3a4452">31</text><text x="689.5" y="258.0" font-size="9.5" text-anchor="middle" fill="#3a4452" font-weight="600">NVS-Solver</text><rect x="46.0" y="278.0" width="11" height="11" fill="#1f77b4" opacity="0.9"/><text x="61.0" y="287.0" font-size="9.5" fill="#54606e">DL3DV (static-real)</text><rect x="182.4" y="278.0" width="11" height="11" fill="#ff7f0e" opacity="0.9"/><text x="197.4" y="287.0" font-size="9.5" fill="#54606e">Kubric (synthetic)</text><rect x="313.2" y="278.0" width="11" height="11" fill="#2ca02c" opacity="0.9"/><text x="328.2" y="287.0" font-size="9.5" fill="#54606e">SpatialVID (dyn-real)</text></svg></div>
<div class="tblwrap"><table><thead><tr><th class="lt" rowspan="2">Method</th><th class="grp" colspan="3">Return-PSNR degrade&darr; (dB, lower = better memory)</th><th class="lt" rowspan="2">memory mechanism &rarr; verdict</th></tr><tr><th>DL3DV<br><span class="nbadge">static-real</span></th><th>Kubric-4D<br><span class="nbadge">synthetic</span></th><th>SpatialVID<br><span class="nbadge">dyn-real</span></th></tr></thead><tbody>
<tr><td class="lt mname">Shape-of-Motion · 4D-GS <span class="psbadge">per-scene 4D-GS</span></td><td style="color:#0a7d2c;font-weight:700">&minus;0.83</td><td style="color:#b6bdc6">&mdash;</td><td style="color:#0a7d2c;font-weight:700">&minus;1.01</td><td class="lt" style="font-size:.74rem"><strong>geometric &mdash; stores the scene</strong></td></tr>
<tr><td class="lt mname">GEN3C · Cosmos-7B <span class="psbadge">3D point cache</span></td><td style="color:#b06f00;font-weight:700">&minus;17.04</td><td style="color:#b06f00;font-weight:700">&minus;13.85</td><td style="color:#b06f00;font-weight:700">&minus;17.52</td><td class="lt" style="font-size:.74rem"><strong>explicit 3D cache &mdash; best generator</strong></td></tr>
<tr class="ours"><td class="lt mname">Ours · VACE-14B <span class="psbadge">warp re-derive</span></td><td style="color:#9a2b15;font-weight:600">&minus;28.18</td><td style="color:#9a2b15;font-weight:600">&minus;21.65</td><td style="color:#9a2b15;font-weight:600">&minus;29.21</td><td class="lt" style="font-size:.74rem">memoryless (re-derives from warp)</td></tr>
<tr><td class="lt mname">Lyra 2.0 · Wan2.1-14B <span class="psbadge">single-shot</span></td><td style="color:#9a2b15;font-weight:600">&minus;28.12</td><td style="color:#9a2b15;font-weight:600">&minus;19.37</td><td style="color:#9a2b15;font-weight:600">&minus;27.62</td><td class="lt" style="font-size:.74rem">memoryless</td></tr>
<tr><td class="lt mname">spmem · spatial-memory <span class="psbadge">learned memory</span></td><td style="color:#9a2b15;font-weight:600">&minus;29.99</td><td style="color:#9a2b15;font-weight:600">&minus;24.08</td><td style="color:#9a2b15;font-weight:600">&minus;32.00</td><td class="lt" style="font-size:.74rem"><b>learned memory &mdash; collapses too</b></td></tr>
<tr><td class="lt mname">NVS-Solver <span class="psbadge">single-shot</span></td><td style="color:#9a2b15;font-weight:600">&minus;31.93</td><td style="color:#9a2b15;font-weight:600">&minus;21.03</td><td style="color:#9a2b15;font-weight:600">&minus;30.97</td><td class="lt" style="font-size:.74rem">memoryless</td></tr>
</tbody></table></div>
<div class="note">Shape-of-Motion is a <i>per-scene 4D-GS reconstruction</i> (the no-forgetting upper bound, not a
forward generator) and is N/A on Kubric-4D where the OURS row is the warp-anchored VACE generator. Read across the
three columns: the only method that holds on <b>all three</b> scene types is GEN3C&rsquo;s explicit 3D cache; the
dedicated learned-memory model <b>spmem collapses hardest of the generators on every type</b> (&minus;24 to
&minus;32&nbsp;dB), confirming the finding is about <i>caching mechanism</i>, not domain or motion.</div>

<h3>M1 &middot; multi_visit &mdash; revisit-consistency</h3>
<div class="sub">Revisit a fixed <i>observed</i> pose 3&times; over the trajectory. <b>&#9733; Revisit
GT-fidelity</b> = PSNR/SSIM/LPIPS vs the real GT frame at that pose (does the revisit match reality?).
<b>Cross-visit drift</b> = self-PSNR/SSIM/LPIPS <i>between</i> the method's own visits to the same pose
(GT-free &mdash; does it stay consistent with itself?). High self-PSNR with low GT-PSNR = consistently
wrong; per-scene 4D-GS shows near-perfect self-consistency (it renders one fixed scene) as the reference.</div>
<h4 style="margin:1.1em 0 .1em;font-size:1.0rem;color:#0b2545">M1a &middot; Kubric-4D clean-GT <span class="tag pri">PRIMARY</span> &mdash; revisit a fixed observed pose (no warp confound)</h4><div class="tblwrap"><table><thead><tr><th class="lt" rowspan="2">Method</th><th class="grp star" colspan="3">&#9733; Revisit GT-fidelity (vs real GT at revisited pose)</th><th class="grp" colspan="3">Cross-visit drift (GT-free self-consistency)</th><th rowspan="2">n</th></tr><tr><th class="star">PSNR&uarr;</th><th class="star">SSIM&uarr;</th><th class="star">LPIPS&darr;</th><th>self-PSNR&uarr;</th><th>self-SSIM&uarr;</th><th>self-LPIPS&darr;</th></tr></thead><tbody><tr class="ours"><td class="lt mname">Ours · Wan2.1-VACE-14B <span class="tag ours">OURS</span></td><td>19.46</td><td>0.772</td><td><strong>0.229</strong></td><td>33.69</td><td>0.955</td><td>0.026</td><td class="nbadge">30</td></tr>
<tr><td class="lt mname">GEN3C · Cosmos-7B</td><td>20.41</td><td><strong>0.832</strong></td><td>0.246</td><td>25.68</td><td>0.888</td><td>0.122</td><td class="nbadge">30</td></tr>
<tr><td class="lt mname">ViewCrafter</td><td>13.58</td><td>0.471</td><td>0.584</td><td>14.24</td><td>0.513</td><td>0.500</td><td class="nbadge">30</td></tr>
<tr><td class="lt mname">FrameCrafter · Wan2.1-14B</td><td>16.33</td><td>0.417</td><td>0.459</td><td>18.88</td><td>0.612</td><td>0.386</td><td class="nbadge">27</td></tr>
<tr><td class="lt mname">NVS-Solver</td><td><strong>20.41</strong></td><td>0.762</td><td>0.241</td><td>28.09</td><td>0.896</td><td>0.081</td><td class="nbadge">30</td></tr>
<tr><td class="lt mname">Lyra 2.0 · Wan2.1-14B</td><td>19.39</td><td>0.787</td><td>0.252</td><td>26.91</td><td>0.899</td><td>0.082</td><td class="nbadge">30</td></tr>
<tr><td class="lt mname">spmem · spatial-memory <span class="tag std">memory</span></td><td>15.87</td><td>0.495</td><td>0.511</td><td>21.05</td><td>0.712</td><td>0.320</td><td class="nbadge">30</td></tr>
<tr><td class="lt mname">Matrix-Game 3.0</td><td>11.10</td><td>0.484</td><td>0.735</td><td>15.89</td><td>0.724</td><td>0.330</td><td class="nbadge">30</td></tr></tbody></table></div><h4 style="margin:1.1em 0 .1em;font-size:1.0rem;color:#0b2545">M1b &middot; DL3DV static &mdash; revisit a fixed observed pose 3&times;</h4><div class="tblwrap"><table><thead><tr><th class="lt" rowspan="2">Method</th><th class="grp star" colspan="3">&#9733; Revisit GT-fidelity (vs real GT at revisited pose)</th><th class="grp" colspan="3">Cross-visit drift (GT-free self-consistency)</th><th rowspan="2">n</th></tr><tr><th class="star">PSNR&uarr;</th><th class="star">SSIM&uarr;</th><th class="star">LPIPS&darr;</th><th>self-PSNR&uarr;</th><th>self-SSIM&uarr;</th><th>self-LPIPS&darr;</th></tr></thead><tbody><tr><td class="lt mname">GEN3C · Cosmos-7B</td><td>15.86</td><td>0.484</td><td>0.485</td><td>15.61</td><td>0.477</td><td>0.469</td><td class="nbadge">14</td></tr>
<tr><td class="lt mname">ViewCrafter</td><td>12.50</td><td>0.323</td><td>0.636</td><td>13.65</td><td>0.379</td><td>0.535</td><td class="nbadge">14</td></tr>
<tr><td class="lt mname">FrameCrafter · Wan2.1-14B</td><td>16.69</td><td>0.481</td><td>0.389</td><td>16.99</td><td>0.519</td><td>0.381</td><td class="nbadge">12</td></tr>
<tr><td class="lt mname">NVS-Solver</td><td>11.78</td><td>0.349</td><td>0.726</td><td>12.82</td><td>0.437</td><td>0.618</td><td class="nbadge">14</td></tr>
<tr><td class="lt mname">Shape-of-Motion · 4D-GS <span class="psbadge" title="per-scene 4D-GS reconstruction (fits the held-out frames) — a reconstruction upper bound, NOT a forward generative pass like the others">per-scene 4D-GS</span></td><td><strong>19.59</strong></td><td><strong>0.628</strong></td><td><strong>0.479</strong></td><td>57.81</td><td>0.998</td><td>0.001</td><td class="nbadge">14</td></tr>
<tr><td class="lt mname">Lyra 2.0 · Wan2.1-14B</td><td>13.42</td><td>0.385</td><td>0.643</td><td>14.92</td><td>0.462</td><td>0.564</td><td class="nbadge">14</td></tr>
<tr><td class="lt mname">spmem · spatial-memory <span class="tag std">memory</span></td><td>11.99</td><td>0.351</td><td>0.730</td><td>14.87</td><td>0.483</td><td>0.527</td><td class="nbadge">12</td></tr>
<tr><td class="lt mname">Matrix-Game 3.0</td><td>10.81</td><td>0.353</td><td>0.761</td><td>13.16</td><td>0.491</td><td>0.607</td><td class="nbadge">14</td></tr></tbody></table></div><h4 style="margin:1.1em 0 .1em;font-size:1.0rem;color:#0b2545">M1c &middot; SpatialVID in-the-wild dynamic &mdash; revisit</h4><div class="tblwrap"><table><thead><tr><th class="lt" rowspan="2">Method</th><th class="grp star" colspan="3">&#9733; Revisit GT-fidelity (vs real GT at revisited pose)</th><th class="grp" colspan="3">Cross-visit drift (GT-free self-consistency)</th><th rowspan="2">n</th></tr><tr><th class="star">PSNR&uarr;</th><th class="star">SSIM&uarr;</th><th class="star">LPIPS&darr;</th><th>self-PSNR&uarr;</th><th>self-SSIM&uarr;</th><th>self-LPIPS&darr;</th></tr></thead><tbody><tr class="ours"><td class="lt mname">Ours · Wan2.1-VACE-14B <span class="tag ours">OURS</span></td><td>15.33</td><td>0.513</td><td>0.507</td><td>14.71</td><td>0.498</td><td>0.525</td><td class="nbadge">27</td></tr>
<tr><td class="lt mname">GEN3C · Cosmos-7B</td><td>16.86</td><td>0.567</td><td><strong>0.406</strong></td><td>17.59</td><td>0.591</td><td>0.378</td><td class="nbadge">30</td></tr>
<tr><td class="lt mname">ViewCrafter</td><td>9.10</td><td>0.258</td><td>0.798</td><td>9.86</td><td>0.257</td><td>0.733</td><td class="nbadge">29</td></tr>
<tr><td class="lt mname">FrameCrafter · Wan2.1-14B</td><td>17.60</td><td>0.533</td><td>0.362</td><td>18.51</td><td>0.611</td><td>0.340</td><td class="nbadge">28</td></tr>
<tr><td class="lt mname">NVS-Solver</td><td>12.83</td><td>0.441</td><td>0.634</td><td>15.52</td><td>0.603</td><td>0.468</td><td class="nbadge">30</td></tr>
<tr><td class="lt mname">Shape-of-Motion · 4D-GS <span class="psbadge" title="per-scene 4D-GS reconstruction (fits the held-out frames) — a reconstruction upper bound, NOT a forward generative pass like the others">per-scene 4D-GS</span></td><td><strong>17.54</strong></td><td><strong>0.598</strong></td><td>0.557</td><td>52.83</td><td>0.996</td><td>0.002</td><td class="nbadge">28</td></tr>
<tr><td class="lt mname">Lyra 2.0 · Wan2.1-14B</td><td>15.42</td><td>0.495</td><td>0.441</td><td>18.15</td><td>0.578</td><td>0.289</td><td class="nbadge">30</td></tr>
<tr><td class="lt mname">spmem · spatial-memory <span class="tag std">memory</span></td><td>11.18</td><td>0.437</td><td>0.689</td><td>13.35</td><td>0.611</td><td>0.523</td><td class="nbadge">27</td></tr>
<tr><td class="lt mname">Matrix-Game 3.0</td><td>10.13</td><td>0.391</td><td>0.721</td><td>13.22</td><td>0.525</td><td>0.556</td><td class="nbadge">30</td></tr></tbody></table></div>
<div class="note">Self-consistency (cross-visit) is trivially near-perfect for the per-scene 4D-GS
reconstructor (it renders ONE optimized scene every visit &mdash; self-PSNR ~50&ndash;58&nbsp;dB) &mdash;
that is the geometric-memory <i>upper bound</i>, not a generative result. Among forward generators, GT-fidelity
on Kubric-4D clean-GT separates the memory-equipped (Lyra&nbsp;2.0 19.4&nbsp;dB) from the memoryless
(Matrix-Game&nbsp;3.0 11.1&nbsp;dB). Ours(base VACE) is competitive on the primary clean-GT revisit
(19.5&nbsp;dB) by re-deriving the pose from warp conditioning.</div>

<h3>M2 &middot; moving_back &mdash; return-fidelity</h3>
<div class="sub">Move away into unseen space, then <b>return</b> to a previously-observed pose. <b>&#9733;
Return fidelity</b> = PSNR/SSIM/LPIPS at the returned pose vs GT. <b>obs-PSNR</b> = the method's own quality
at that pose <i>before</i> departing (the baseline). <b>degrade</b> = the PSNR drop after the round-trip
(<b>the memory/forgetting metric</b> &mdash; smaller is better). <b>away</b> = mean camera distance travelled
before returning.</div>
<h4 style="margin:1.1em 0 .1em;font-size:1.0rem;color:#0b2545">M2a &middot; Kubric-4D clean-GT <span class="tag pri">PRIMARY</span> &mdash; move away into unseen space, then return</h4><div class="tblwrap"><table><thead><tr><th class="lt" rowspan="2">Method</th><th class="grp star" colspan="3">&#9733; Return fidelity (back at the observed pose)</th><th class="grp" colspan="3">Return context</th><th rowspan="2">n</th></tr><tr><th class="star">PSNR&uarr;</th><th class="star">SSIM&uarr;</th><th class="star">LPIPS&darr;</th><th title="observed-pose PSNR baseline before moving away">obs-PSNR&uarr;</th><th title="PSNR drop vs the observed baseline after the round-trip — the memory/forgetting signal (smaller = better)">degrade&darr;</th><th title="mean camera distance travelled away before returning">away</th></tr></thead><tbody><tr class="ours"><td class="lt mname">Ours · Wan2.1-VACE-14B <span class="tag ours">OURS</span></td><td>18.42</td><td>0.723</td><td>0.274</td><td>40.07</td><td>+21.65</td><td>19.92</td><td class="nbadge">30</td></tr>
<tr><td class="lt mname">GEN3C · Cosmos-7B</td><td><strong>22.44</strong></td><td><strong>0.879</strong></td><td><strong>0.185</strong></td><td>36.28</td><td>+13.85</td><td>19.92</td><td class="nbadge">30</td></tr>
<tr><td class="lt mname">ViewCrafter</td><td>13.64</td><td>0.472</td><td>0.573</td><td>39.00</td><td>+25.35</td><td>19.92</td><td class="nbadge">30</td></tr>
<tr><td class="lt mname">FrameCrafter · Wan2.1-14B</td><td>16.08</td><td>0.376</td><td>0.513</td><td>40.07</td><td>+24.00</td><td>19.92</td><td class="nbadge">30</td></tr>
<tr><td class="lt mname">NVS-Solver</td><td>19.05</td><td>0.769</td><td>0.243</td><td>40.08</td><td>+21.03</td><td>19.86</td><td class="nbadge">24</td></tr>
<tr><td class="lt mname">Lyra 2.0 · Wan2.1-14B</td><td>20.69</td><td>0.836</td><td>0.199</td><td>40.06</td><td>+19.37</td><td>19.57</td><td class="nbadge">29</td></tr>
<tr><td class="lt mname">spmem · spatial-memory <span class="tag std">memory</span></td><td>16.00</td><td>0.496</td><td>0.514</td><td>40.08</td><td>+24.08</td><td>19.92</td><td class="nbadge">30</td></tr>
<tr><td class="lt mname">Matrix-Game 3.0</td><td>11.06</td><td>0.494</td><td>0.752</td><td>20.61</td><td><strong>+9.55</strong></td><td>19.92</td><td class="nbadge">30</td></tr></tbody></table></div><h4 style="margin:1.1em 0 .1em;font-size:1.0rem;color:#0b2545">M2b &middot; DL3DV static &mdash; move away, then return to the observed pose</h4><div class="tblwrap"><table><thead><tr><th class="lt" rowspan="2">Method</th><th class="grp star" colspan="3">&#9733; Return fidelity (back at the observed pose)</th><th class="grp" colspan="3">Return context</th><th rowspan="2">n</th></tr><tr><th class="star">PSNR&uarr;</th><th class="star">SSIM&uarr;</th><th class="star">LPIPS&darr;</th><th title="observed-pose PSNR baseline before moving away">obs-PSNR&uarr;</th><th title="PSNR drop vs the observed baseline after the round-trip — the memory/forgetting signal (smaller = better)">degrade&darr;</th><th title="mean camera distance travelled away before returning">away</th></tr></thead><tbody><tr><td class="lt mname">GEN3C · Cosmos-7B</td><td>17.30</td><td>0.519</td><td><strong>0.399</strong></td><td>34.33</td><td>+17.04</td><td>4.66</td><td class="nbadge">14</td></tr>
<tr><td class="lt mname">ViewCrafter</td><td>12.26</td><td>0.312</td><td>0.659</td><td>40.71</td><td>+28.45</td><td>4.66</td><td class="nbadge">14</td></tr>
<tr><td class="lt mname">FrameCrafter · Wan2.1-14B</td><td>10.72</td><td>0.244</td><td>0.730</td><td>41.72</td><td>+31.00</td><td>4.81</td><td class="nbadge">12</td></tr>
<tr><td class="lt mname">NVS-Solver</td><td>9.88</td><td>0.301</td><td>0.787</td><td>41.82</td><td>+31.93</td><td>4.66</td><td class="nbadge">14</td></tr>
<tr><td class="lt mname">Shape-of-Motion · 4D-GS <span class="psbadge" title="per-scene 4D-GS reconstruction (fits the held-out frames) — a reconstruction upper bound, NOT a forward generative pass like the others">per-scene 4D-GS</span></td><td><strong>20.02</strong></td><td><strong>0.642</strong></td><td>0.442</td><td>20.86</td><td><strong>+0.83</strong></td><td>4.66</td><td class="nbadge">14</td></tr>
<tr><td class="lt mname">Lyra 2.0 · Wan2.1-14B</td><td>13.69</td><td>0.379</td><td>0.633</td><td>41.82</td><td>+28.12</td><td>4.66</td><td class="nbadge">14</td></tr>
<tr><td class="lt mname">spmem · spatial-memory <span class="tag std">memory</span></td><td>11.73</td><td>0.344</td><td>0.738</td><td>41.72</td><td>+29.99</td><td>4.81</td><td class="nbadge">12</td></tr>
<tr><td class="lt mname">Matrix-Game 3.0</td><td>11.25</td><td>0.343</td><td>0.732</td><td>13.06</td><td>+1.82</td><td>4.66</td><td class="nbadge">14</td></tr></tbody></table></div><h4 style="margin:1.1em 0 .1em;font-size:1.0rem;color:#0b2545">M2c &middot; SpatialVID in-the-wild dynamic &mdash; move-back</h4><div class="tblwrap"><table><thead><tr><th class="lt" rowspan="2">Method</th><th class="grp star" colspan="3">&#9733; Return fidelity (back at the observed pose)</th><th class="grp" colspan="3">Return context</th><th rowspan="2">n</th></tr><tr><th class="star">PSNR&uarr;</th><th class="star">SSIM&uarr;</th><th class="star">LPIPS&darr;</th><th title="observed-pose PSNR baseline before moving away">obs-PSNR&uarr;</th><th title="PSNR drop vs the observed baseline after the round-trip — the memory/forgetting signal (smaller = better)">degrade&darr;</th><th title="mean camera distance travelled away before returning">away</th></tr></thead><tbody><tr class="ours"><td class="lt mname">Ours · Wan2.1-VACE-14B <span class="tag ours">OURS</span></td><td>13.03</td><td>0.440</td><td>0.616</td><td>42.24</td><td>+29.21</td><td>5.25</td><td class="nbadge">27</td></tr>
<tr><td class="lt mname">GEN3C · Cosmos-7B</td><td><strong>17.30</strong></td><td>0.581</td><td><strong>0.365</strong></td><td>34.81</td><td>+17.52</td><td>5.16</td><td class="nbadge">30</td></tr>
<tr><td class="lt mname">ViewCrafter</td><td>9.07</td><td>0.262</td><td>0.776</td><td>41.60</td><td>+32.54</td><td>5.18</td><td class="nbadge">29</td></tr>
<tr><td class="lt mname">FrameCrafter · Wan2.1-14B</td><td>9.63</td><td>0.214</td><td>0.764</td><td>42.19</td><td>+32.56</td><td>5.16</td><td class="nbadge">30</td></tr>
<tr><td class="lt mname">NVS-Solver</td><td>11.22</td><td>0.370</td><td>0.706</td><td>42.19</td><td>+30.97</td><td>5.16</td><td class="nbadge">30</td></tr>
<tr><td class="lt mname">Shape-of-Motion · 4D-GS <span class="psbadge" title="per-scene 4D-GS reconstruction (fits the held-out frames) — a reconstruction upper bound, NOT a forward generative pass like the others">per-scene 4D-GS</span></td><td>17.00</td><td><strong>0.583</strong></td><td>0.569</td><td>18.00</td><td><strong>+1.01</strong></td><td>5.25</td><td class="nbadge">28</td></tr>
<tr><td class="lt mname">Lyra 2.0 · Wan2.1-14B</td><td>14.57</td><td>0.475</td><td>0.521</td><td>42.20</td><td>+27.62</td><td>5.16</td><td class="nbadge">30</td></tr>
<tr><td class="lt mname">spmem · spatial-memory <span class="tag std">memory</span></td><td>10.24</td><td>0.409</td><td>0.743</td><td>42.24</td><td>+32.00</td><td>5.25</td><td class="nbadge">27</td></tr>
<tr><td class="lt mname">Matrix-Game 3.0</td><td>11.33</td><td>0.376</td><td>0.620</td><td>15.01</td><td>+3.68</td><td>5.16</td><td class="nbadge">30</td></tr></tbody></table></div>
<h4 style="margin:1.1em 0 .1em;font-size:1.0rem;color:#0b2545">M3 &middot; Kubric-4D return-fidelity by away-distance bin <span class="tag pri">PRIMARY</span> <span class="tag std">clean-GT</span></h4><div class="sub">The angle bin = how far the camera travelled into unseen space before returning (30&deg; nearby &rarr; 180&deg; full back-and-return). Return-PSNR holds roughly flat across bins for the warp-anchored methods (they re-derive the observed region from the conditioning), exposing whether quality is driven by <i>memory</i> or by re-warping.</div><div class="tblwrap"><table><thead><tr><th class="lt" rowspan="2">Method</th><th class="grp" colspan="2">30&deg;</th><th class="grp" colspan="2">60&deg;</th><th class="grp" colspan="2">90&deg;</th><th class="grp" colspan="2">120&deg;</th><th class="grp" colspan="2">180&deg;</th></tr><tr><th class="bin">ret-PSNR&uarr;</th><th>degrade&darr;</th><th class="bin">ret-PSNR&uarr;</th><th>degrade&darr;</th><th class="bin">ret-PSNR&uarr;</th><th>degrade&darr;</th><th class="bin">ret-PSNR&uarr;</th><th>degrade&darr;</th><th class="bin">ret-PSNR&uarr;</th><th>degrade&darr;</th></tr></thead><tbody><tr class="ours"><td class="lt mname">Ours · Wan2.1-VACE-14B <span class="tag ours">OURS</span></td><td>18.36</td><td>+21.7</td><td>18.49</td><td>+21.6</td><td>18.42</td><td>+21.7</td><td>18.40</td><td>+21.7</td><td>18.44</td><td>+21.6</td></tr>
<tr><td class="lt mname">GEN3C · Cosmos-7B</td><td>22.62</td><td>+13.7</td><td>22.49</td><td>+13.8</td><td>22.57</td><td>+13.7</td><td>22.31</td><td>+14.0</td><td>22.19</td><td>+14.1</td></tr>
<tr><td class="lt mname">ViewCrafter</td><td>13.64</td><td>+25.4</td><td>13.64</td><td>+25.4</td><td>13.64</td><td>+25.4</td><td>13.64</td><td>+25.4</td><td>13.64</td><td>+25.4</td></tr>
<tr><td class="lt mname">FrameCrafter · Wan2.1-14B</td><td>16.08</td><td>+24.0</td><td>16.08</td><td>+24.0</td><td>16.08</td><td>+24.0</td><td>16.08</td><td>+24.0</td><td>16.08</td><td>+24.0</td></tr>
<tr><td class="lt mname">NVS-Solver</td><td>18.75</td><td>+21.3</td><td>18.95</td><td>+21.1</td><td>19.64</td><td>+20.3</td><td>19.07</td><td>+21.0</td><td>18.96</td><td>+21.1</td></tr>
<tr><td class="lt mname">Lyra 2.0 · Wan2.1-14B</td><td>20.43</td><td>+19.7</td><td>20.58</td><td>+19.5</td><td>20.59</td><td>+19.5</td><td>20.56</td><td>+19.5</td><td>21.39</td><td>+18.6</td></tr>
<tr><td class="lt mname">Matrix-Game 3.0</td><td>11.06</td><td>+9.5</td><td>11.06</td><td>+9.5</td><td>11.06</td><td>+9.5</td><td>11.06</td><td>+9.5</td><td>11.06</td><td>+9.5</td></tr></tbody></table></div>
<div class="note warn">&#9888; <b>How to read &lsquo;degrade&rsquo;:</b> a large positive degrade means the
method was sharp at the pose before leaving but <b>forgot it</b> after wandering &mdash; pure generative
re-hallucination, no memory. Per-scene 4D-GS (~&minus;0.8 to &minus;1.0&nbsp;dB) is the no-forgetting
reference (it stores the whole scene); the explicit 3D-cache generator GEN3C (&minus;17.0 / &minus;13.8&nbsp;dB)
is the best forward method, while the single-shot generators&rsquo; &minus;20&nbsp;to&nbsp;&minus;32&nbsp;dB
collapse is the open problem this axis exposes. <b>Crucially, the dedicated learned spatial-memory model
spmem collapses just as hard (&minus;30.0 DL3DV / &minus;24.1 Kubric / &minus;32.0 SpatialVID&nbsp;dB) on all three
scene types</b> &mdash; learned memory is not yet recalling the observed pose; explicit geometric caching is. <b>Honest scope:</b> these benches
are young &mdash; 9 methods scored per source now (spmem and FrameCrafter both folded in across <b>all six</b> memory benchmarks,
confirming the collapse on static-real, synthetic AND dynamic-real; FlexWorld/TrajCrafter/ViewExtrapolator/Voyager
and spmem on the still-generating long-horizon set remain <span class="run">pending</span>);
the numbers shown are real-GT means, never fabricated.</div>
</section>
</section>