| <section id="story"> |
| <style> |
| #story{font-family:-apple-system,BlinkMacSystemFont,"Segoe UI",Roboto,Helvetica,Arial,sans-serif;color:#1b1f24;max-width:100%;margin:0 auto;padding:8px 4px 30px;} |
| #story h2{font-size:1.55rem;margin:.2em 0 .2em;font-weight:750;letter-spacing:-.01em;} |
| #story .lede{font-size:.96rem;line-height:1.6;color:#3a4452;max-width:66rem;margin:.2em 0 .3em;} |
| #story .lede b{color:#0b2545;} |
| #story .lede .ours{color:#0a7d2c;font-weight:700;} |
| #story .subnote{font-size:.78rem;color:#6b7480;margin:.1em 0 1.1em;line-height:1.5;} |
| #story .grid{display:grid;grid-template-columns:repeat(3,minmax(0,1fr));gap:14px;margin:.4em 0 .2em;} |
| #story .card{border:1px solid #e3e8ef;border-radius:10px;background:#fff;box-shadow:0 1px 2px rgba(20,30,55,.04);display:flex;flex-direction:column;overflow:hidden;} |
| #story .card .hd{padding:9px 13px 7px;border-bottom:1px solid #eef1f5;} |
| #story .card .axis{font-size:.7rem;font-weight:800;letter-spacing:.06em;text-transform:uppercase;color:#8a929c;} |
| #story .card .ttl{font-size:1.02rem;font-weight:750;color:#0b2545;margin:.1em 0 0;letter-spacing:-.01em;} |
| #story .card .bench{font-size:.74rem;color:#6b7480;margin:.15em 0 0;} |
| #story .card.static .hd{background:linear-gradient(180deg,#f0f7f1,#fff);} |
| #story .card.dynamic .hd{background:linear-gradient(180deg,#eef3fb,#fff);} |
| #story .card.longh .hd{background:linear-gradient(180deg,#faf3ec,#fff);} |
| #story .card.static{border-top:3px solid #0a7d2c;} |
| #story .card.dynamic{border-top:3px solid #15539e;} |
| #story .card.longh{border-top:3px solid #e0a800;} |
| #story ul.find{list-style:none;margin:0;padding:9px 13px 11px;font-size:.83rem;line-height:1.5;color:#3a4452;flex:1;} |
| #story ul.find li{position:relative;padding:0 0 0 16px;margin:0 0 7px;} |
| #story ul.find li:last-child{margin-bottom:0;} |
| #story ul.find li::before{content:"";position:absolute;left:0;top:.5em;width:6px;height:6px;border-radius:50%;} |
| #story .card.static ul.find li::before{background:#0a7d2c;} |
| #story .card.dynamic ul.find li::before{background:#15539e;} |
| #story .card.longh ul.find li::before{background:#e0a800;} |
| #story ul.find b{color:#0b2545;font-weight:700;} |
| #story ul.find .num{font-variant-numeric:tabular-nums;font-weight:700;color:#0a7d2c;} |
| #story ul.find .warn{color:#9a2b15;font-weight:700;} |
| #story .tradeoff{margin:1.3em 0 .3em;border:1px solid #f0d9b0;border-left:5px solid #e0922f;border-radius:0 9px 9px 0;background:#fff8ee;padding:12px 16px;} |
| #story .tradeoff .th{font-size:.78rem;font-weight:800;letter-spacing:.06em;text-transform:uppercase;color:#9a6a16;margin:0 0 .4em;} |
| #story .tradeoff ul{list-style:none;margin:0;padding:0;font-size:.86rem;line-height:1.55;color:#5a4a2c;} |
| #story .tradeoff li{position:relative;padding:0 0 0 18px;margin:0 0 6px;} |
| #story .tradeoff li:last-child{margin-bottom:0;} |
| #story .tradeoff li::before{content:"\26A0";position:absolute;left:0;top:0;font-size:.8rem;color:#e0922f;} |
| #story .tradeoff li.plus::before{content:"\2713";color:#0a7d2c;} |
| #story .tradeoff b{color:#4a3a1c;} |
| #story .tradeoff .ratio{font-variant-numeric:tabular-nums;font-weight:800;color:#9a2b15;} |
| #story .bottom{margin:1.1em 0 0;font-size:.95rem;line-height:1.55;color:#102a16;background:#eef7ef;border:1px solid #cfe6d2;border-radius:8px;padding:11px 15px;} |
| #story .bottom b{color:#0a4d1c;} |
| #story .bottom .lab{font-weight:800;letter-spacing:.04em;text-transform:uppercase;font-size:.72rem;color:#0a7d2c;margin-right:8px;} |
| #story .honest{font-size:.74rem;color:#8a929c;margin:.9em 0 0;line-height:1.5;} |
| @media (max-width:820px){#story .grid{grid-template-columns:1fr;}} |
| </style> |
|
|
| <h2>Executive summary & key findings</h2> |
| <p class="lede">Per-scene 3D Gaussian Splatting is superb at interpolating densely observed captures — |
| and brittle exactly where it is asked to invent: <b>extrapolative</b> held-out views with big |
| disocclusion holes, <b>dynamic</b> moving content, and <b>long-horizon</b> chained / loop-closing |
| trajectories that accumulate drift. Diffusion-based generative NVS replaces 3D-GS in precisely those |
| regimes. <span class="ours">Our method — Wan2.1-VACE-14B warp+inpaint</span> warps the input to each |
| target pose, then lets video diffusion fill the disocclusions <b>hole-free</b> while staying |
| <b>geometrically anchored</b> to the warped scaffold.</p> |
| <p class="subnote">Every number below is read verbatim off the scored runs (eval50, Kubric-4D clean-GT, |
| NVIDIA, SpatialVID, LH-DL3DV, WorldScore) and SLURM <code>sacct</code> timing — nothing is |
| interpolated. "Ours" = <code>vace14b</code>; best-in-column values are noted as such.</p> |
|
|
| <div class="grid"> |
|
|
| <div class="card static"> |
| <div class="hd"> |
| <div class="axis">Axis 1 · Static</div> |
| <div class="ttl">Extrapolation, hole-free</div> |
| <div class="bench">DL3DV eval50 · 50 held-out clips</div> |
| </div> |
| <ul class="find"> |
| <li>Warp+inpaint yields geometry-consistent extrapolation with <b>no torn disocclusions</b>: Ours leads generation realism at <span class="num">FID 56.4</span> (best in column).</li> |
| <li>Tied at the front on perceptual fidelity over the <b>hole region only</b> — PSNR-extrap <span class="num">12.79</span> (vs GEN3C 12.78), full-frame LPIPS <span class="num">0.534</span> (tied-best).</li> |
| <li>Strongest pose anchoring: COLMAP <b>registration rate 0.761</b> (best in column) — the warped scaffold keeps generated frames self-consistent.</li> |
| <li>All methods cluster low on raw PSNR (this is <b>generation, not reconstruction</b>); the real separation lives in FID + geometric consistency, where Ours leads.</li> |
| </ul> |
| </div> |
|
|
| <div class="card dynamic"> |
| <div class="hd"> |
| <div class="axis">Axis 2 · Dynamic</div> |
| <div class="ttl">Movers are hard for all</div> |
| <div class="bench">Kubric-4D clean-GT (primary) · NVIDIA · SpatialVID</div> |
| </div> |
| <ul class="find"> |
| <li><b>Every method degrades as the viewpoint widens 30°→180°.</b> Honest moving-object-masked mPSNR <span class="warn">collapses to ~11–14 dB</span> across all bins for everyone — the static background hides this in full-frame PSNR.</li> |
| <li>Ours shows the strongest <b>pose-free multi-view consistency</b>: MEt3R <span class="num">0.014</span> on Kubric-4D (best; next-best 0.034) and <span class="num">0.027</span> on NVIDIA (best, with TSED 0.80, rot 1.0°).</li> |
| <li>Camera-trajectory error stays <b>low and flat</b> across difficulty bins (COLMAP ATE <span class="num">0.024→0.050</span> scene-units, 30°→180°).</li> |
| <li>Ours is absent from the pixel-PSNR-vs-angle curves <b>by design</b>: it does not follow the exact GT poses, so reconstruction metrics don't apply — it is judged on pose-free geometry/realism instead.</li> |
| </ul> |
| </div> |
|
|
| <div class="card longh"> |
| <div class="hd"> |
| <div class="axis">Axis 3 · Long-horizon</div> |
| <div class="ttl">On-trajectory under drift</div> |
| <div class="bench">LH-DL3DV drift/loop · WorldScore (ICCV'25)</div> |
| </div> |
| <ul class="find"> |
| <li>Warp-anchored control stays on-trajectory across long rollouts: drift <b>ATE 3.48%</b> of scene extent, rotational drift <span class="num">2.5°</span> — while single-shot baselines <span class="warn">collapse rotationally (52–116°)</span>.</li> |
| <li>Best long-horizon realism: <span class="num">FID 50.9</span> (best in column) over 228-frame rollouts.</li> |
| <li>On <b>WorldScore</b>, Ours is <b>2nd overall (67.8)</b>, best photometric quality (88.9) and 2nd on camera controllability (78.3); NVS-Solver leads overall (70.4).</li> |
| </ul> |
| </div> |
|
|
| </div> |
|
|
| <div class="tradeoff"> |
| <div class="th">Honest tradeoffs</div> |
| <ul> |
| <li class="plus"><b>What Ours wins:</b> hole-free extrapolated content, best-in-class geometric / multi-view consistency, and 2nd place on the WorldScore composite — without any per-scene optimization.</li> |
| <li><b>Runtime is mid-pack, not fastest.</b> Feed-forward world models are far cheaper per clip: on long-horizon, Ours runs <span class="ratio">2732 s/clip</span> vs Lyra-2's 464 (<span class="ratio">5.9×</span> slower) and MatrixGame-3's 354 (<span class="ratio">7.7×</span> slower). The geometric-consistency win costs roughly 6–8× the inference budget of the fastest feed-forward baselines.</li> |
| <li><b>Dynamic movers remain unsolved for everyone.</b> Masked mPSNR pins at ~11–14 dB across all methods and all angle bins on clean-GT Kubric-4D — reconstructing fast-moving content from a single view is an open problem, not a per-method gap.</li> |
| </ul> |
| </div> |
|
|
| <p class="bottom"><span class="lab">Bottom line</span> |
| Warp+inpaint VACE-14B delivers <b>hole-free, geometry-anchored novel views that stay on-trajectory</b> |
| where 3D-GS fails — best FID and consistency, 2nd on WorldScore — at a mid-pack inference cost |
| (~6–8× the fastest feed-forward models), with dynamic movers still hard for the whole field.</p> |
|
|
| <p class="honest">Sources: <code>quantitative.html</code> tables, <code>runs/dyn_kubric4d_metrics/*.json</code> |
| (per-method aggregate + by-bin), and <code>infer_time.json</code> (SLURM <code>sacct</code> s/clip). Ratios |
| computed from long-horizon secs_per_clip. Where Ours has no value for a metric (e.g. exact-pose |
| reconstruction PSNR), it is stated as not-applicable rather than estimated.</p> |
| </section> |
|
|
