Add Avg inference-time column (SLURM sacct) to quantitative tables + §5 runtime summary; infer_time.json
11cbba0 verified | <section id="quantitative"> | |
| <style> | |
| #quantitative{font-family:-apple-system,BlinkMacSystemFont,"Segoe UI",Roboto,Helvetica,Arial,sans-serif;color:#1b1f24;max-width:100%;margin:0 auto;padding:8px 4px 40px;} | |
| #quantitative h2{font-size:1.55rem;margin:.2em 0 .15em;font-weight:700;letter-spacing:-.01em;} | |
| #quantitative h3{font-size:1.12rem;margin:1.4em 0 .25em;font-weight:650;color:#0b2545;border-bottom:2px solid #e3e8ef;padding-bottom:3px;} | |
| #quantitative .sub{font-size:.86rem;color:#5b6470;margin:.1em 0 .7em;line-height:1.4;} | |
| #quantitative .tblwrap{overflow-x:auto;border:1px solid #e3e8ef;border-radius:7px;margin:.3em 0 .4em;} | |
| #quantitative table{border-collapse:collapse;width:100%;font-size:.8rem;font-variant-numeric:tabular-nums;} | |
| #quantitative 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;} | |
| #quantitative thead th.lt{text-align:left;} | |
| #quantitative 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;} | |
| #quantitative thead th.star{background:#7a4f00;} | |
| #quantitative tbody td{padding:5px 8px;text-align:right;white-space:nowrap;border-bottom:1px solid #eef1f5;} | |
| #quantitative tbody td.lt{text-align:left;} | |
| #quantitative tbody tr:nth-child(even){background:#fafbfc;} | |
| #quantitative tbody tr.ours{background:#fff6e0;} | |
| #quantitative tbody tr.ours:nth-child(even){background:#fdf0d2;} | |
| #quantitative tbody tr.ours td{border-bottom:1px solid #f0dca8;} | |
| #quantitative tbody tr.oursft{background:#f3f7ff;} | |
| #quantitative tbody tr.oursft:nth-child(even){background:#ecf2fc;} | |
| #quantitative tbody tr:hover{background:#eef4fb;} | |
| #quantitative tbody td strong{color:#0a7d2c;font-weight:750;} | |
| #quantitative .na{color:#b6bdc6;} | |
| #quantitative .run{color:#b06f00;font-style:italic;font-size:.74rem;} | |
| #quantitative .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;} | |
| #quantitative .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;} | |
| #quantitative .tag.ours{background:#f0a500;color:#231a00;} | |
| #quantitative .tag.pri{background:#0a7d2c;color:#fff;} | |
| #quantitative .tag.sec{background:#9aa3ad;color:#fff;} | |
| #quantitative .tag.std{background:#15539e;color:#fff;} | |
| #quantitative .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;} | |
| #quantitative .note.warn{background:#fff7ed;border-left-color:#e0922f;color:#6b4a16;} | |
| #quantitative .note b{color:#1b1f24;} | |
| #quantitative .legend{font-size:.74rem;color:#6b7480;margin:.2em 0 1em;} | |
| #quantitative .legend code{background:#eef1f5;border-radius:3px;padding:0 4px;} | |
| #quantitative .mname{font-size:.78rem;} | |
| #quantitative td.bin{border-left:1px solid #e3e8ef;} | |
| #quantitative .nbadge{color:#7a828c;font-weight:400;font-size:.72rem;} | |
| #quantitative th.infer{background:#1d3b1d;} | |
| #quantitative td.infer{text-align:right;color:#155e2b;font-weight:600;} | |
| #quantitative td.infer .spf{color:#7a9a82;font-weight:400;font-size:.72rem;} | |
| #quantitative td.infer.slow{color:#9a2b15;background:#fdeee8;} | |
| #quantitative td.infer.na{color:#b6bdc6;font-weight:400;} | |
| #quantitative td .psbadge{display:inline-block;background:#eef3f7;color:#5a6b78;border-radius:4px;padding:0 5px;font-size:.66rem;font-weight:600;vertical-align:middle;} | |
| #quantitative td .spf{color:#7a828c;font-weight:400;font-size:.72rem;} | |
| #quantitative thead th .u{font-weight:400;font-size:.66rem;opacity:.85;} | |
| </style> | |
| <h2>Quantitative Benchmarks</h2> | |
| <div class="legend">Arrows show metric polarity: <b>↑</b> higher-is-better, <b>↓</b> lower-is-better. | |
| <strong style="color:#0a7d2c">Bold green</strong> = best in column. <span style="background:#fff6e0;padding:0 5px;border-radius:3px">Highlighted row</span> = Ours (VACE-14B). | |
| <code>—</code> = no value; <span class="run">running</span> = method has outputs but the metric is not yet computed. Every cell is read verbatim from <code>bench_data.json</code> — no value is invented.</div> | |
| <h3>1 · Static novel-view extrapolation — DL3DV (50 held-out clips)</h3> | |
| <div class="sub">PSNR / SSIM / LPIPS are scored over the <b>hole / extrapolated region only</b> (eval_metrics <code>psnr_extrap</code> etc.) — the unobserved pixels a method must hallucinate. Full-frame variants shown as secondary where available. Geometry: TSED (SfM 2-view inlier frac), MEt3R (multi-view consistency), COLMAP registration rate & rotation error.</div> | |
| <div class="tblwrap"><table> | |
| <thead> | |
| <tr> | |
| <th class="lt" rowspan="2">Method</th> | |
| <th class="grp" colspan="3">Hole / extrapolated region (primary)</th> | |
| <th class="grp" colspan="3">Full frame (secondary)</th> | |
| <th class="grp" colspan="2">Generation realism</th> | |
| <th class="grp" colspan="3">Geometry / consistency</th> | |
| <th rowspan="2">n</th> | |
| <th rowspan="2" class="infer" title="avg wall-clock seconds per clip on 1xH200, incl. model-load amortized over the generation array (sacct ElapsedRaw / clips produced)">infer s/clip↓</th></tr> | |
| <tr> | |
| <th>PSNR↑</th><th>SSIM↑</th><th>LPIPS↓</th> | |
| <th>PSNR↑</th><th>SSIM↑</th><th>LPIPS↓</th> | |
| <th>FID↓</th><th>TSED↑</th> | |
| <th>MEt3R↓</th><th>reg↑</th><th>rot°↓</th> | |
| </tr> | |
| </thead><tbody> | |
| <tr class="ours"><td class="lt mname">Ours · Wan2.1-VACE-14B <span class="tag ours">OURS</span></td><td>12.79</td><td>0.374</td><td>0.605</td><td>14.06</td><td>0.411</td><td>0.534</td><td><strong>56.4</strong></td><td>0.498</td><td>0.135</td><td>0.761</td><td>17.7</td><td class="nbadge">50</td><td class="infer" title="758 s/clip = 9.36 s/frame x 81f; 50 clips, 79 sacct tasks">758<span class="spf"> (9.36/f)</span></td></tr> | |
| <tr><td class="lt mname">GEN3C · Cosmos-7B</td><td>12.78</td><td><strong>0.385</strong></td><td>0.661</td><td>14.25</td><td>0.425</td><td>0.605</td><td><span class="run">running</span></td><td>0.258</td><td>0.135</td><td>0.302</td><td>68.6</td><td class="nbadge">50</td><td class="infer" title="498 s/clip = 6.14 s/frame x 81f; 50 clips, 10 sacct tasks">498<span class="spf"> (6.14/f)</span></td></tr> | |
| <tr><td class="lt mname">Voyager · HunyuanVideo</td><td>11.05</td><td>0.338</td><td>0.741</td><td>11.32</td><td>0.350</td><td>0.725</td><td>136.3</td><td>0.179</td><td>0.109</td><td>0.341</td><td>93.7</td><td class="nbadge">49</td><td class="infer" title="1086 s/clip = 13.41 s/frame x 81f; 49 clips, 147 sacct tasks">1086<span class="spf"> (13.41/f)</span></td></tr> | |
| <tr><td class="lt mname">FlexWorld</td><td>12.88</td><td>0.373</td><td>0.657</td><td>13.54</td><td>0.389</td><td>0.613</td><td>67.2</td><td>0.266</td><td>0.121</td><td>0.631</td><td>24.1</td><td class="nbadge">50</td><td class="infer" title="259 s/clip = 3.20 s/frame x 81f; 50 clips, 1 sacct tasks">259<span class="spf"> (3.20/f)</span></td></tr> | |
| <tr><td class="lt mname">ViewCrafter</td><td>11.83</td><td>0.318</td><td>0.627</td><td><span class="na">—</span></td><td><span class="na">—</span></td><td><span class="na">—</span></td><td><span class="run">running</span></td><td>0.582</td><td>0.144</td><td>0.654</td><td><strong>4.2</strong></td><td class="nbadge">36</td><td class="infer" title="264 s/clip = 3.26 s/frame x 81f; 50 clips, 10 sacct tasks">264<span class="spf"> (3.26/f)</span></td></tr> | |
| <tr><td class="lt mname">FrameCrafter · Wan2.1-14B</td><td><strong>14.38</strong></td><td>0.377</td><td><strong>0.448</strong></td><td><span class="na">—</span></td><td><span class="na">—</span></td><td><span class="na">—</span></td><td><span class="run">running</span></td><td><strong>0.639</strong></td><td>0.150</td><td>0.756</td><td>16.1</td><td class="nbadge">50</td><td class="infer slow" title="2413 s/clip = 29.80 s/frame x 81f; 50 clips, 10 sacct tasks">2413<span class="spf"> (29.80/f)</span></td></tr> | |
| <tr><td class="lt mname">NVS-Solver</td><td>12.98</td><td>0.380</td><td>0.641</td><td><span class="na">—</span></td><td><span class="na">—</span></td><td><span class="na">—</span></td><td><span class="run">running</span></td><td>0.483</td><td><strong>0.088</strong></td><td><strong>0.776</strong></td><td>43.9</td><td class="nbadge">45</td><td class="infer" title="122 s/clip = 1.51 s/frame x 81f; 50 clips, 1 sacct tasks">122<span class="spf"> (1.51/f)</span></td></tr> | |
| <tr><td class="lt mname">ViewExtrapolator</td><td>11.23</td><td>0.291</td><td>0.742</td><td>12.96</td><td>0.341</td><td>0.675</td><td>116.6</td><td>0.205</td><td>0.221</td><td>0.269</td><td>80.1</td><td class="nbadge">50</td><td class="infer" title="131 s/clip = 1.61 s/frame x 81f; 50 clips, 1 sacct tasks">131<span class="spf"> (1.61/f)</span></td></tr> | |
| <tr><td class="lt mname">TrajectoryCrafter</td><td><span class="run">running</span></td><td><span class="run">running</span></td><td><span class="run">running</span></td><td><span class="na">—</span></td><td><span class="na">—</span></td><td><span class="na">—</span></td><td>87.1</td><td>0.108</td><td>0.193</td><td>0.481</td><td>134.7</td><td class="nbadge">0</td><td class="infer" title="610 s/clip = 12.46 s/frame x 49f; 48 clips, 79 sacct tasks">610<span class="spf"> (12.46/f)</span></td></tr> | |
| </tbody></table></div> | |
| <div class="note">Full-frame PSNR/SSIM/LPIPS are present only for methods whose harness emitted them (vace14b, gen3c, flexworld, voyager, viewextrapolator); others show <code>—</code>. All methods cluster low on raw PSNR (this is <i>generation</i>, not reconstruction); separation lives in perceptual LPIPS, FID and geometric consistency. Ours leads <b>FID (56.4)</b> and is essentially tied-best on full-frame LPIPS (0.535).</div> | |
| <h3>2 · Dynamic-scene novel-view synthesis</h3> | |
| <div class="sub">The <b>confound-free primary anchor is Kubric-4D clean-GT</b> (perfect GT poses/depth/segmentation — the headline <a href="#difficulty" style="color:#0a7d2c;font-weight:600">angle-binned difficulty curve</a> is up top). <b>NVIDIA</b> (dynamic-region masked) is a <b>secondary</b> real-dynamic check — its 12-camera static rig leaks multi-view cues (the “teleporting camera”). DyCheck (iPhone portrait) was dropped: its portrait aspect ratio doesn't fit the landscape pipeline.</div> | |
| <h4 style="margin:1.1em 0 .1em;font-size:1.0rem;color:#0b2545">2a · NVIDIA Dynamic Scenes <span class="tag sec">SECONDARY</span></h4> | |
| <div class="sub">24 clips, 12-camera static rig. Masked = dynamic-region.</div> | |
| <div class="tblwrap"><table> | |
| <thead><tr><th class="lt" rowspan="2">Method</th><th class="grp star" colspan="3">★ Masked — dynamic-region (primary)</th><th class="grp" colspan="3">Full frame (secondary)</th><th class="grp" colspan="2">Realism</th><th class="grp" colspan="1">Consist.</th><th rowspan="2">n</th><th rowspan="2">m·n</th><th rowspan="2" class="infer" title="avg wall-clock seconds per clip on 1xH200, incl. model-load amortized over the generation array (sacct ElapsedRaw / clips produced)">infer s/clip↓</th></tr><tr><th class="star">mPSNR↑</th><th class="star">mSSIM↑</th><th class="star">mLPIPS↓</th><th>PSNR↑</th><th>SSIM↑</th><th>LPIPS↓</th><th>FID↓</th><th>TSED↑</th><th>MEt3R↓</th></tr></thead><tbody> | |
| <tr class="ours"><td class="lt mname">Ours · Wan2.1-VACE-14B <span class="tag ours">OURS</span></td><td><span class="run">running</span></td><td><span class="run">running</span></td><td><span class="run">running</span></td><td><span class="run">running</span></td><td><span class="run">running</span></td><td><span class="run">running</span></td><td>205.3</td><td><strong>0.800</strong></td><td><strong>0.027</strong></td><td class="nbadge">0</td><td class="nbadge">0</td><td class="infer" title="795 s/clip = 9.81 s/frame x 81f; 18 clips, 6 sacct tasks">795<span class="spf"> (9.81/f)</span></td></tr> | |
| <tr><td class="lt mname">GEN3C · Cosmos-7B</td><td>13.80</td><td><strong>0.338</strong></td><td>0.069</td><td><strong>19.53</strong></td><td><strong>0.641</strong></td><td><strong>0.275</strong></td><td>71.1</td><td>0.139</td><td>0.102</td><td class="nbadge">21</td><td class="nbadge">21</td><td class="infer" title="779 s/clip = 32.46 s/frame x 24f; 21 clips, 6 sacct tasks">779<span class="spf"> (32.46/f)</span></td></tr> | |
| <tr><td class="lt mname">Voyager · HunyuanVideo</td><td>11.39</td><td>0.236</td><td>0.081</td><td>14.32</td><td>0.296</td><td>0.539</td><td>187.6</td><td>0.039</td><td>0.071</td><td class="nbadge">18</td><td class="nbadge">18</td><td class="infer slow" title="2395 s/clip = 99.79 s/frame x 24f; 18 clips, 6 sacct tasks">2395<span class="spf"> (99.79/f)</span></td></tr> | |
| <tr><td class="lt mname">FlexWorld</td><td>12.00</td><td>0.251</td><td>0.076</td><td>14.73</td><td>0.345</td><td>0.470</td><td>184.7</td><td>0.673</td><td>0.067</td><td class="nbadge">18</td><td class="nbadge">18</td><td class="infer" title="663 s/clip = 27.64 s/frame x 24f; 18 clips, 6 sacct tasks">663<span class="spf"> (27.64/f)</span></td></tr> | |
| <tr><td class="lt mname">ViewCrafter</td><td>12.26</td><td>0.262</td><td>0.090</td><td>12.14</td><td>0.299</td><td>0.701</td><td>91.2</td><td>0.263</td><td>0.107</td><td class="nbadge">21</td><td class="nbadge">21</td><td class="infer" title="505 s/clip = 21.03 s/frame x 24f; 21 clips, 6 sacct tasks">505<span class="spf"> (21.03/f)</span></td></tr> | |
| <tr><td class="lt mname">FrameCrafter · Wan2.1-14B</td><td><strong>13.96</strong></td><td>0.270</td><td><strong>0.067</strong></td><td>17.50</td><td>0.378</td><td>0.334</td><td><strong>69.7</strong></td><td>0.296</td><td>0.124</td><td class="nbadge">21</td><td class="nbadge">21</td><td class="infer" title="742 s/clip = 30.89 s/frame x 24f; 21 clips, 3 sacct tasks">742<span class="spf"> (30.89/f)</span></td></tr> | |
| <tr><td class="lt mname">NVS-Solver</td><td>13.85</td><td>0.335</td><td>0.076</td><td>16.71</td><td>0.392</td><td>0.557</td><td>82.8</td><td>0.232</td><td>0.106</td><td class="nbadge">18</td><td class="nbadge">18</td><td class="infer" title="318 s/clip = 13.24 s/frame x 24f; 18 clips, 6 sacct tasks">318<span class="spf"> (13.24/f)</span></td></tr> | |
| <tr><td class="lt mname">ViewExtrapolator</td><td>12.32</td><td>0.263</td><td>0.079</td><td>15.28</td><td>0.340</td><td>0.538</td><td>213.9</td><td>0.114</td><td>0.124</td><td class="nbadge">18</td><td class="nbadge">18</td><td class="infer" title="144 s/clip = 5.99 s/frame x 24f; 18 clips, 9 sacct tasks">144<span class="spf"> (5.99/f)</span></td></tr> | |
| <tr><td class="lt mname">TrajectoryCrafter</td><td><span class="run">running</span></td><td><span class="run">running</span></td><td><span class="run">running</span></td><td><span class="run">running</span></td><td><span class="run">running</span></td><td><span class="run">running</span></td><td>104.8</td><td>0.079</td><td>0.083</td><td class="nbadge">0</td><td class="nbadge">0</td><td class="infer" title="514 s/clip = 10.50 s/frame x 49f; 14 clips, 12 sacct tasks">514<span class="spf"> (10.50/f)</span></td></tr> | |
| <tr><td class="lt mname">Shape-of-Motion · 4D-GS</td><td>11.24</td><td>0.296</td><td>0.096</td><td>13.26</td><td>0.412</td><td>0.796</td><td>355.2</td><td>0.228</td><td>0.292</td><td class="nbadge">21</td><td class="nbadge">21</td><td class="infer na">—</td></tr> | |
| </tbody></table></div> | |
| <div class="note warn">⚠ <b>NVIDIA Dynamic Scenes uses a 12-camera static rig</b> (high effective multi-view factor — the “teleporting camera” flagged by DyCheck’s <i>Reality Check</i>). Treat as <b>SECONDARY</b>; the confound-free primary dynamic anchor is <b>Kubric-4D clean-GT</b> (perfect ground-truth, no warp confound — see the difficulty curve at the top). Note <b>vace14b currently has n=0 / null masked recon</b> on NVIDIA (warp caveat) — only geometry metrics (TSED 0.80, rot 1.0°, MEt3R 0.027, all best-in-col) are populated; its recon cells render as <span class="run">running</span>.</div> | |
| <h4 style="margin:1.1em 0 .1em;font-size:1.0rem;color:#0b2545">2b · Kubric-4D (synthetic dynamic — summary) <span class="tag pri">PRIMARY</span> <span class="tag std">STANDARD GCD/AnyView</span></h4> | |
| <div class="sub">Aggregate across all angle bins (100 clips total). PERFECT-GT — <b>no depth/scale/warp confound (the confound-free primary dynamic anchor)</b>. The headline <a href="#difficulty" style="color:#0a7d2c;font-weight:600">angle-binned difficulty curve</a> is at the top of the page; the per-bin table is Section 3 below. Ours runs are still being scored (recon n=0 → PSNR/SSIM/LPIPS show <span class="run">running</span>).</div> | |
| <div class="tblwrap"><table> | |
| <thead><tr><th class="lt">Method</th><th>PSNR↑</th><th>SSIM↑</th><th>LPIPS↓</th><th>FID↓</th><th>TSED↑</th><th>MEt3R↓</th><th>reg↑</th><th>rot°↓</th><th>n</th><th class="infer" title="avg wall-clock seconds per clip on 1xH200, incl. model-load amortized over the generation array (sacct ElapsedRaw / clips produced)">infer s/clip↓</th></tr></thead><tbody> | |
| <tr class="ours"><td class="lt mname">Ours · Wan2.1-VACE-14B <span class="tag ours">OURS</span></td><td><span class="run">running</span></td><td><span class="run">running</span></td><td><span class="run">running</span></td><td><span class="run">running</span></td><td>0.094</td><td><strong>0.014</strong></td><td>0.145</td><td>144.8</td><td class="nbadge">0</td><td class="infer" title="289 s/clip = 3.57 s/frame x 81f; 100 clips, 18 sacct tasks">289<span class="spf"> (3.57/f)</span></td></tr> | |
| <tr><td class="lt mname">GEN3C · Cosmos-7B</td><td>13.68</td><td><strong>0.540</strong></td><td>0.537</td><td><strong>71.5</strong></td><td>0.101</td><td>0.034</td><td>0.192</td><td>120.1</td><td class="nbadge">100</td><td class="infer" title="503 s/clip = 20.96 s/frame x 24f; 100 clips, 15 sacct tasks">503<span class="spf"> (20.96/f)</span></td></tr> | |
| <tr><td class="lt mname">Voyager · HunyuanVideo</td><td>11.89</td><td>0.489</td><td>0.568</td><td>241.2</td><td>0.085</td><td>0.056</td><td><strong>0.758</strong></td><td><strong>113.0</strong></td><td class="nbadge">5</td><td class="infer slow" title="1901 s/clip = 79.22 s/frame x 24f; 5 clips, 18 sacct tasks">1901<span class="spf"> (79.22/f)</span></td></tr> | |
| <tr><td class="lt mname">FlexWorld</td><td>12.47</td><td>0.422</td><td>0.563</td><td>171.8</td><td><strong>0.109</strong></td><td>0.042</td><td>0.350</td><td>132.6</td><td class="nbadge">100</td><td class="infer" title="459 s/clip = 19.12 s/frame x 24f; 100 clips, 18 sacct tasks">459<span class="spf"> (19.12/f)</span></td></tr> | |
| <tr><td class="lt mname">ViewCrafter</td><td>12.34</td><td>0.505</td><td>0.609</td><td>78.0</td><td>0.025</td><td>0.110</td><td>0.363</td><td>146.8</td><td class="nbadge">100</td><td class="infer" title="226 s/clip = 9.41 s/frame x 24f; 100 clips, 18 sacct tasks">226<span class="spf"> (9.41/f)</span></td></tr> | |
| <tr><td class="lt mname">FrameCrafter · Wan2.1-14B</td><td>12.44</td><td>0.371</td><td>0.631</td><td>107.5</td><td>0.018</td><td>0.114</td><td>0.204</td><td>141.6</td><td class="nbadge">100</td><td class="infer" title="866 s/clip = 36.08 s/frame x 24f; 100 clips, 12 sacct tasks">866<span class="spf"> (36.08/f)</span></td></tr> | |
| <tr><td class="lt mname">NVS-Solver</td><td><strong>14.52</strong></td><td>0.479</td><td><strong>0.522</strong></td><td>81.8</td><td>0.065</td><td>0.063</td><td>0.580</td><td>135.0</td><td class="nbadge">100</td><td class="infer" title="230 s/clip = 9.59 s/frame x 24f; 100 clips, 18 sacct tasks">230<span class="spf"> (9.59/f)</span></td></tr> | |
| <tr><td class="lt mname">TrajectoryCrafter</td><td><span class="run">running</span></td><td><span class="run">running</span></td><td><span class="run">running</span></td><td>123.9</td><td>0.052</td><td>0.089</td><td>0.579</td><td>141.9</td><td class="nbadge">0</td><td class="infer" title="336 s/clip = 6.86 s/frame x 49f; 67 clips, 18 sacct tasks">336<span class="spf"> (6.86/f)</span></td></tr> | |
| <tr><td class="lt mname">Shape-of-Motion · 4D-GS</td><td>5.38</td><td>0.370</td><td>0.855</td><td>336.8</td><td>0.035</td><td>0.064</td><td>0.017</td><td>131.2</td><td class="nbadge">15</td><td class="infer" title="1487 s/clip = 61.98 s/frame x 24f; 15 clips, 214 sacct tasks; per-scene optim">1487<span class="spf"> (61.98/f ·ps)</span></td></tr> | |
| </tbody></table></div> | |
| <h4 style="margin:1.1em 0 .1em;font-size:1.0rem;color:#0b2545">2c · SpatialVID (in-the-wild dynamic)</h4> | |
| <div class="sub">30 clips (group_0001).</div> | |
| <div class="tblwrap"><table> | |
| <thead><tr><th class="lt">Method</th><th>PSNR↑</th><th>SSIM↑</th><th>LPIPS↓</th><th>FID↓</th><th>TSED↑</th><th>MEt3R↓</th><th>reg↑</th><th>rot°↓</th><th>n</th><th class="infer" title="avg wall-clock seconds per clip on 1xH200, incl. model-load amortized over the generation array (sacct ElapsedRaw / clips produced)">infer s/clip↓</th></tr></thead><tbody> | |
| <tr class="ours"><td class="lt mname">Ours · Wan2.1-VACE-14B <span class="tag ours">OURS</span></td><td>14.20</td><td>0.349</td><td>0.669</td><td>86.2</td><td><strong>0.897</strong></td><td>0.072</td><td>0.909</td><td><strong>23.2</strong></td><td class="nbadge">16</td><td class="infer" title="389 s/clip = 4.81 s/frame x 81f; 30 clips, 6 sacct tasks">389<span class="spf"> (4.81/f)</span></td></tr> | |
| <tr><td class="lt mname">GEN3C · Cosmos-7B</td><td>14.91</td><td><strong>0.536</strong></td><td>0.479</td><td><span class="run">running</span></td><td>0.765</td><td><strong>0.058</strong></td><td>0.771</td><td>89.3</td><td class="nbadge">20</td><td class="infer" title="455 s/clip = 5.68 s/frame x 80f; 30 clips, 6 sacct tasks">455<span class="spf"> (5.68/f)</span></td></tr> | |
| <tr><td class="lt mname">Voyager · HunyuanVideo</td><td>12.57</td><td>0.425</td><td>0.593</td><td>69.7</td><td>0.666</td><td>0.065</td><td>0.914</td><td>76.5</td><td class="nbadge">30</td><td class="infer" title="1103 s/clip = 13.79 s/frame x 80f; 30 clips, 6 sacct tasks">1103<span class="spf"> (13.79/f)</span></td></tr> | |
| <tr><td class="lt mname">ViewCrafter</td><td>11.43</td><td>0.346</td><td>0.711</td><td><span class="run">running</span></td><td>0.647</td><td>0.100</td><td>0.466</td><td>47.1</td><td class="nbadge">29</td><td class="infer" title="299 s/clip = 3.73 s/frame x 80f; 30 clips, 6 sacct tasks">299<span class="spf"> (3.73/f)</span></td></tr> | |
| <tr><td class="lt mname">FrameCrafter · Wan2.1-14B</td><td><strong>14.99</strong></td><td>0.442</td><td><strong>0.475</strong></td><td><span class="run">running</span></td><td>0.633</td><td>0.138</td><td>0.807</td><td>102.7</td><td class="nbadge">27</td><td class="infer slow" title="2262 s/clip = 28.27 s/frame x 80f; 30 clips, 6 sacct tasks">2262<span class="spf"> (28.27/f)</span></td></tr> | |
| <tr><td class="lt mname">NVS-Solver</td><td>14.34</td><td>0.462</td><td>0.561</td><td><strong>33.5</strong></td><td>0.664</td><td>0.079</td><td>0.637</td><td>69.3</td><td class="nbadge">30</td><td class="infer" title="150 s/clip = 1.88 s/frame x 80f; 30 clips, 6 sacct tasks">150<span class="spf"> (1.88/f)</span></td></tr> | |
| <tr><td class="lt mname">ViewExtrapolator</td><td>13.02</td><td>0.390</td><td>0.641</td><td>85.9</td><td>0.674</td><td>0.111</td><td>0.625</td><td>65.3</td><td class="nbadge">30</td><td class="infer" title="129 s/clip = 1.61 s/frame x 80f; 30 clips, 6 sacct tasks">129<span class="spf"> (1.61/f)</span></td></tr> | |
| <tr><td class="lt mname">TrajectoryCrafter</td><td><span class="run">running</span></td><td><span class="run">running</span></td><td><span class="run">running</span></td><td>53.3</td><td>0.517</td><td>0.090</td><td><strong>0.958</strong></td><td>114.7</td><td class="nbadge">0</td><td class="infer" title="284 s/clip = 5.80 s/frame x 49f; 24 clips, 6 sacct tasks">284<span class="spf"> (5.80/f)</span></td></tr> | |
| </tbody></table></div> | |
| <div class="note">Kubric-4D & SpatialVID use full-frame PSNR/SSIM/LPIPS as the primary recon metrics (plus moving-object mask on Kubric). On SpatialVID, Ours is best on <b>TSED (0.897)</b> and rotation error (23.2°). On Kubric-4D summary, Ours recon is still <span class="run">running</span>; its MEt3R (0.014) is already best-in-column.</div> | |
| <h3>5 · Average inference time — wall-clock per clip (SLURM accounting)</h3> | |
| <div class="sub">Honest <b>wall-clock seconds per clip on 1×H200</b>, from <code>sacct</code> | |
| <code>ElapsedRaw</code> summed over the <b>COMPLETED generation array tasks</b> divided by the clips that | |
| method produced on that benchmark (NOT mtime deltas, which parallel array writes make meaningless). | |
| Each cell: <b>s/clip</b> with <b>(s/frame)</b> in parentheses. Sorted fastest→slowest by mean | |
| s/frame — surfacing the spread from fast feed-forward (<b>Lyra 2.0 / Matrix-Game</b>, | |
| ~1.5–2 s/frame) to autoregressive-memory (<b>Voyager</b>, 13–100 s/frame) to the | |
| per-scene <b>Shape-of-Motion</b> 4D-GS optimization (~62 s/frame, full per-scene fit). | |
| <code>—</code> = no sacct generation record in the 14-day window (never | |
| fabricated).</div> | |
| <div class="tblwrap"><table> | |
| <thead><tr><th class="lt">Method</th><th>DL3DV static<br><span class="u">s/clip (s/frame)</span></th><th>Kubric-4D<br><span class="u">s/clip (s/frame)</span></th><th>NVIDIA dyn<br><span class="u">s/clip (s/frame)</span></th><th>SpatialVID<br><span class="u">s/clip (s/frame)</span></th><th>Long-horizon<br><span class="u">s/clip (s/frame)</span></th></tr></thead><tbody> | |
| <tr><td class="lt mname">Matrix-Game 3.0</td><td class="na">—</td><td class="na">—</td><td class="na">—</td><td class="na">—</td><td>354<span class="spf"> (1.55)</span></td></tr> | |
| <tr><td class="lt mname">Lyra 2.0 · Wan2.1-14B</td><td class="na">—</td><td class="na">—</td><td class="na">—</td><td class="na">—</td><td>464<span class="spf"> (2.03)</span></td></tr> | |
| <tr><td class="lt mname">ViewExtrapolator</td><td>131<span class="spf"> (1.61)</span></td><td class="na">—</td><td>144<span class="spf"> (5.99)</span></td><td>129<span class="spf"> (1.61)</span></td><td>690<span class="spf"> (3.03)</span></td></tr> | |
| <tr><td class="lt mname">NVS-Solver</td><td>122<span class="spf"> (1.51)</span></td><td>230<span class="spf"> (9.59)</span></td><td>318<span class="spf"> (13.24)</span></td><td>150<span class="spf"> (1.88)</span></td><td>2288<span class="spf"> (10.04)</span></td></tr> | |
| <tr><td class="lt mname">TrajectoryCrafter</td><td>610<span class="spf"> (12.46)</span></td><td>336<span class="spf"> (6.86)</span></td><td>514<span class="spf"> (10.50)</span></td><td>284<span class="spf"> (5.80)</span></td><td>828<span class="spf"> (3.63)</span></td></tr> | |
| <tr class="ours"><td class="lt mname">Ours · Wan2.1-VACE-14B <span class="tag ours">OURS</span></td><td>758<span class="spf"> (9.36)</span></td><td>289<span class="spf"> (3.57)</span></td><td>795<span class="spf"> (9.81)</span></td><td>389<span class="spf"> (4.81)</span></td><td>2732<span class="spf"> (11.98)</span></td></tr> | |
| <tr><td class="lt mname">ViewCrafter</td><td>264<span class="spf"> (3.26)</span></td><td>226<span class="spf"> (9.41)</span></td><td>505<span class="spf"> (21.03)</span></td><td>299<span class="spf"> (3.73)</span></td><td>843<span class="spf"> (3.70)</span></td></tr> | |
| <tr><td class="lt mname">I2V base · Wan2.1-14B</td><td>670<span class="spf"> (8.28)</span></td><td class="na">—</td><td class="na">—</td><td class="na">—</td><td class="na">—</td></tr> | |
| <tr><td class="lt mname">FlexWorld</td><td>259<span class="spf"> (3.20)</span></td><td>459<span class="spf"> (19.12)</span></td><td>663<span class="spf"> (27.64)</span></td><td class="na">—</td><td>1095<span class="spf"> (4.80)</span></td></tr> | |
| <tr><td class="lt mname">GEN3C · Cosmos-7B</td><td>498<span class="spf"> (6.14)</span></td><td>503<span class="spf"> (20.96)</span></td><td>779<span class="spf"> (32.46)</span></td><td>455<span class="spf"> (5.68)</span></td><td>1148<span class="spf"> (5.04)</span></td></tr> | |
| <tr class="oursft"><td class="lt mname">Ours-ft · VACE-14B (long-horizon FT)</td><td class="na">—</td><td class="na">—</td><td class="na">—</td><td class="na">—</td><td>3253<span class="spf"> (14.27)</span></td></tr> | |
| <tr><td class="lt mname">Long-term Spatial Memory</td><td class="na">—</td><td class="na">—</td><td class="na">—</td><td class="na">—</td><td>5929<span class="spf"> (26.00)</span></td></tr> | |
| <tr><td class="lt mname">FrameCrafter · Wan2.1-14B</td><td>2413<span class="spf"> (29.80)</span></td><td>866<span class="spf"> (36.08)</span></td><td>742<span class="spf"> (30.89)</span></td><td>2262<span class="spf"> (28.27)</span></td><td>7296<span class="spf"> (32.00)</span></td></tr> | |
| <tr><td class="lt mname">Voyager · HunyuanVideo</td><td>1086<span class="spf"> (13.41)</span></td><td>1901<span class="spf"> (79.22)</span></td><td>2395<span class="spf"> (99.79)</span></td><td>1103<span class="spf"> (13.79)</span></td><td>4914<span class="spf"> (21.55)</span></td></tr> | |
| <tr><td class="lt mname">Shape-of-Motion · 4D-GS <span class="psbadge" title="per-scene 4D-GS optimization, not a forward pass">per-scene</span></td><td class="na">—</td><td>1487<span class="spf"> (61.98)</span></td><td class="na">—</td><td class="na">—</td><td class="na">—</td></tr> | |
| </tbody></table></div> | |
| <div class="note"><b>Caveats (read inline):</b><ul style="margin:.3em 0 0 1.1em;padding:0"><li>Wall-clock SECONDS PER CLIP on 1xH200, INCLUDING one-time model-load amortized over the generation array (not a pure forward pass).</li><li>Clip LENGTHS differ per benchmark (static/kubric/nvidia/spatialvid ~81 frames; long-horizon ~150-240) -> s/frame is the cross-benchmark-fair number; the per-method output frame count is shown.</li><li>Source = SLURM sacct ElapsedRaw summed over COMPLETED generation array tasks / clips produced. Methods with no sacct record in the last 14 days (purged) show n/a -- never fabricated.</li><li>Per-scene methods (Shape-of-Motion) include FULL per-scene 4D-GS optimization, not just a forward pass.</li></ul> | |
| <span style="display:block;margin-top:.4em">Generated 2026-06-28T21:49:45 from SLURM sacct | |
| (<code>code/compute_infer_time.py</code> → <code>infer_time.json</code>); per-method clip counts & | |
| sacct task counts are in the JSON.</span></div> | |
| <h3>3 · Kubric-4D angle-binned difficulty curve — per-bin table</h3> | |
| <div class="sub">The numbers behind the <a href="#difficulty" style="color:#0a7d2c;font-weight:600">headline difficulty curve</a> at the top. Standard GCD/AnyView protocol, binned by optical-axis change from the input view. Shown per bin: full <b>PSNR↑</b> / <b>LPIPS↓</b>, moving-object-masked <b>mPSNR↑</b>, and flow-warp <b>EPE↓</b> (temporal). Harder bins (120°/180° back-views) = larger viewpoint change. Best per (bin×metric) bolded.</div> | |
| <div class="sub" style="margin-top:-.4em">Clips per bin — 30°: n=20 · 60°: n=20 · 90°: n=20 · 120°: n=20 · 180°: n=20</div> | |
| <div class="tblwrap"><table> | |
| <thead><tr><th class="lt" rowspan="2">Method</th> | |
| <th class="grp" colspan="4">30° <span style="font-weight:400;opacity:.8">(n=20)</span></th> | |
| <th class="grp" colspan="4">60° <span style="font-weight:400;opacity:.8">(n=20)</span></th> | |
| <th class="grp" colspan="4">90° <span style="font-weight:400;opacity:.8">(n=20)</span></th> | |
| <th class="grp" colspan="4">120° <span style="font-weight:400;opacity:.8">(n=20)</span></th> | |
| <th class="grp" colspan="4">180° <span style="font-weight:400;opacity:.8">(n=20)</span></th> | |
| </tr><tr> | |
| <th class="bin">PSNR↑</th><th>LPIPS↓</th><th>mPSNR↑</th><th>EPE↓</th> | |
| <th class="bin">PSNR↑</th><th>LPIPS↓</th><th>mPSNR↑</th><th>EPE↓</th> | |
| <th class="bin">PSNR↑</th><th>LPIPS↓</th><th>mPSNR↑</th><th>EPE↓</th> | |
| <th class="bin">PSNR↑</th><th>LPIPS↓</th><th>mPSNR↑</th><th>EPE↓</th> | |
| <th class="bin">PSNR↑</th><th>LPIPS↓</th><th>mPSNR↑</th><th>EPE↓</th> | |
| </tr></thead><tbody> | |
| <tr><td class="lt mname">GEN3C · Cosmos-7B</td><td class="bin">20.08</td><td>0.323</td><td>11.60</td><td><strong>0.611</strong></td><td class="bin">19.83</td><td>0.337</td><td>11.79</td><td><strong>0.593</strong></td><td class="bin">19.04</td><td>0.355</td><td>11.73</td><td><strong>0.657</strong></td><td class="bin">18.95</td><td>0.366</td><td>11.46</td><td><strong>0.632</strong></td><td class="bin">18.66</td><td>0.360</td><td>11.83</td><td><strong>0.591</strong></td></tr> | |
| <tr><td class="lt mname">Voyager · HunyuanVideo</td><td class="bin">16.30</td><td>0.414</td><td>9.45</td><td>0.815</td><td class="bin">16.48</td><td>0.437</td><td>9.44</td><td>2.687</td><td class="bin">15.33</td><td>0.464</td><td>13.02</td><td>2.550</td><td class="bin">13.46</td><td>0.502</td><td><strong>14.53</strong></td><td>1.730</td><td class="bin">12.97</td><td>0.490</td><td>13.09</td><td>1.764</td></tr> | |
| <tr><td class="lt mname">FlexWorld</td><td class="bin">16.39</td><td>0.333</td><td>13.52</td><td>0.648</td><td class="bin">15.52</td><td>0.385</td><td>12.33</td><td>0.685</td><td class="bin">14.82</td><td>0.441</td><td>11.58</td><td>1.036</td><td class="bin">14.59</td><td>0.452</td><td>11.27</td><td>0.803</td><td class="bin">14.77</td><td>0.433</td><td>11.72</td><td>1.261</td></tr> | |
| <tr><td class="lt mname">ViewCrafter</td><td class="bin">25.30</td><td>0.319</td><td>11.24</td><td>7.147</td><td class="bin">25.19</td><td>0.323</td><td>11.26</td><td>7.147</td><td class="bin">24.92</td><td>0.332</td><td>11.28</td><td>7.147</td><td class="bin">24.63</td><td>0.340</td><td>11.55</td><td>7.147</td><td class="bin">24.59</td><td>0.338</td><td>11.74</td><td>7.147</td></tr> | |
| <tr><td class="lt mname">FrameCrafter · Wan2.1-14B</td><td class="bin"><strong>27.61</strong></td><td>0.267</td><td>11.05</td><td>5.226</td><td class="bin"><strong>27.19</strong></td><td>0.290</td><td>11.29</td><td>5.447</td><td class="bin"><strong>26.62</strong></td><td>0.329</td><td>11.29</td><td>6.074</td><td class="bin"><strong>26.18</strong></td><td>0.353</td><td>10.80</td><td>6.622</td><td class="bin"><strong>25.86</strong></td><td>0.370</td><td>10.17</td><td>7.206</td></tr> | |
| <tr><td class="lt mname">NVS-Solver</td><td class="bin">26.18</td><td><strong>0.258</strong></td><td><strong>14.39</strong></td><td>3.787</td><td class="bin">25.56</td><td><strong>0.273</strong></td><td><strong>14.18</strong></td><td>3.882</td><td class="bin">24.68</td><td><strong>0.293</strong></td><td><strong>13.58</strong></td><td>4.276</td><td class="bin">24.58</td><td><strong>0.296</strong></td><td>13.42</td><td>4.051</td><td class="bin">24.52</td><td><strong>0.291</strong></td><td><strong>13.92</strong></td><td>4.173</td></tr> | |
| <tr><td class="lt mname">TrajectoryCrafter</td><td class="bin"><span class="na">—</span></td><td><span class="na">—</span></td><td><span class="na">—</span></td><td>4.465</td><td class="bin"><span class="na">—</span></td><td><span class="na">—</span></td><td><span class="na">—</span></td><td>5.469</td><td class="bin"><span class="na">—</span></td><td><span class="na">—</span></td><td><span class="na">—</span></td><td>4.971</td><td class="bin"><span class="na">—</span></td><td><span class="na">—</span></td><td><span class="na">—</span></td><td>4.907</td><td class="bin"><span class="na">—</span></td><td><span class="na">—</span></td><td><span class="na">—</span></td><td>4.481</td></tr> | |
| <tr><td class="lt mname">Shape-of-Motion · 4D-GS</td><td class="bin">13.45</td><td>0.559</td><td>8.13</td><td>2.145</td><td class="bin">12.63</td><td>0.591</td><td>6.39</td><td>2.173</td><td class="bin">12.38</td><td>0.606</td><td>6.03</td><td>2.330</td><td class="bin">12.65</td><td>0.588</td><td>5.97</td><td>2.337</td><td class="bin">13.32</td><td>0.570</td><td>6.82</td><td>2.325</td></tr> | |
| </tbody></table></div> | |
| <div class="note"><b>STANDARD GCD/AnyView angle-binned protocol</b>; PERFECT-GT (no depth / scale / warp confound). Harder bins = larger optical-axis change from the input view (120°/180° are back-views). PSNR/LPIPS degrade monotonically with angle for warp-based methods (e.g. ViewCrafter 25.3→~18 dB), while moving-object mPSNR and flow-EPE expose temporal/dynamic fidelity. Ours (vace14b) is not yet scored on the binned recon metrics here (recon n=0) and is omitted from this table.</div> | |
| <h3>4 · Long-horizon geometric stability</h3> | |
| <h4 style="margin:.7em 0 .1em;font-size:1.0rem;color:#0b2545">4a · Drift / loop-closure (GVS-style, DL3DV)</h4> | |
| <div class="sub">Rolled-out long trajectories. ATE & loop-closure error are fractions of scene extent. Sorted by ATE (best first). FID / MEt3R folded in from the lh_dl3dv recon metrics.</div> | |
| <div class="tblwrap"><table> | |
| <thead><tr><th class="lt">Method</th><th>drift ATE↓<br><span style="font-weight:400;font-size:.7rem">(frac extent)</span></th><th>rot drift°↓</th><th>loop-closure↓<br><span style="font-weight:400;font-size:.7rem">(frac extent)</span></th><th>FID↓</th><th>MEt3R↓</th><th>n</th><th class="infer" title="avg wall-clock seconds per clip on 1xH200, incl. model-load amortized over the generation array (sacct ElapsedRaw / clips produced)">infer s/clip↓</th></tr></thead><tbody> | |
| <tr><td class="lt mname">Voyager · HunyuanVideo</td><td><strong>0.03%</strong></td><td><strong>0.5</strong></td><td>0.999</td><td><span class="na">—</span></td><td><strong>0.059</strong></td><td class="nbadge">33</td><td class="infer" title="4914 s/clip = 21.55 s/frame x 228f; 33 clips, 168 sacct tasks">4914<span class="spf"> (21.55/f)</span></td></tr> | |
| <tr><td class="lt mname">GEN3C · Cosmos-7B</td><td>2.40%</td><td>27.3</td><td>0.959</td><td><span class="na">—</span></td><td>0.096</td><td class="nbadge">42</td><td class="infer" title="1148 s/clip = 5.04 s/frame x 228f; 42 clips, 84 sacct tasks">1148<span class="spf"> (5.04/f)</span></td></tr> | |
| <tr><td class="lt mname">FlexWorld</td><td>3.19%</td><td>4.4</td><td>0.925</td><td>86.9</td><td>0.093</td><td class="nbadge">42</td><td class="infer" title="1095 s/clip = 4.80 s/frame x 228f; 42 clips, 126 sacct tasks">1095<span class="spf"> (4.80/f)</span></td></tr> | |
| <tr class="ours"><td class="lt mname">Ours · Wan2.1-VACE-14B <span class="tag ours">OURS</span></td><td>3.48%</td><td>2.5</td><td><span class="na">—</span></td><td><strong>50.9</strong></td><td>0.113</td><td class="nbadge">42</td><td class="infer" title="2732 s/clip = 11.98 s/frame x 228f; 42 clips, 84 sacct tasks">2732<span class="spf"> (11.98/f)</span></td></tr> | |
| <tr><td class="lt mname">ViewCrafter</td><td>6.00%</td><td>24.1</td><td>0.856</td><td>89.4</td><td>0.134</td><td class="nbadge">42</td><td class="infer" title="843 s/clip = 3.70 s/frame x 228f; 42 clips, 84 sacct tasks">843<span class="spf"> (3.70/f)</span></td></tr> | |
| <tr><td class="lt mname">TrajectoryCrafter</td><td>12.64%</td><td>116.2</td><td>0.731</td><td>100.2</td><td>0.090</td><td class="nbadge">42</td><td class="infer" title="828 s/clip = 3.63 s/frame x 228f; 42 clips, 126 sacct tasks">828<span class="spf"> (3.63/f)</span></td></tr> | |
| <tr><td class="lt mname">NVS-Solver</td><td>13.04%</td><td>68.0</td><td><strong>0.407</strong></td><td>86.4</td><td>0.109</td><td class="nbadge">42</td><td class="infer" title="2288 s/clip = 10.04 s/frame x 228f; 42 clips, 84 sacct tasks">2288<span class="spf"> (10.04/f)</span></td></tr> | |
| <tr><td class="lt mname">ViewExtrapolator</td><td>18.40%</td><td>67.5</td><td>0.568</td><td><span class="na">—</span></td><td>0.175</td><td class="nbadge">42</td><td class="infer" title="690 s/clip = 3.03 s/frame x 228f; 42 clips, 84 sacct tasks">690<span class="spf"> (3.03/f)</span></td></tr> | |
| <tr><td class="lt mname">FrameCrafter · Wan2.1-14B</td><td>29.78%</td><td>52.6</td><td><span class="na">—</span></td><td><span class="na">—</span></td><td>0.183</td><td class="nbadge">42</td><td class="infer slow" title="7296 s/clip = 32.00 s/frame x 228f; 42 clips, 84 sacct tasks">7296<span class="spf"> (32.00/f)</span></td></tr> | |
| <tr class="oursft"><td class="lt mname">Ours-ft · VACE-14B (long-horizon FT) <span class="ftlabel">ft@4000 · re-running</span></td><td><span class="run">—</span></td><td><span class="run">—</span></td><td><span class="run">—</span></td><td><span class="run">—</span></td><td><span class="run">—</span></td><td class="nbadge">3</td><td class="infer" title="3253 s/clip = 14.27 s/frame x 228f; 42 clips, 336 sacct tasks">3253<span class="spf"> (14.27/f)</span></td></tr> | |
| </tbody></table></div> | |
| <div class="note">Single-shot SVD methods (nvs_solver / viewextrapolator) and trajcrafter / framecrafter <b>collapse rotationally</b> (52–116° drift); warp-anchored <b>VACE stays geometrically stable</b> (3.5% ATE, 2.5° rot). voyager’s ≈0 ATE comes from near-static rollouts (almost no camera motion). The <code>vace14b_ft_lh</code> long-horizon-finetuned variant (ft@4000) has only <code>.bak</code> outputs — metrics are re-running, shown as em-dashes.</div> | |
| <h4 style="margin:1.0em 0 .1em;font-size:1.0rem;color:#0b2545">4b · WorldScore (ICCV 2025 Stanford world-gen benchmark)</h4> | |
| <div class="sub">Composite world-generation score. Sorted by WorldScore (best first).</div> | |
| <div class="tblwrap"><table> | |
| <thead><tr><th class="lt">Method</th><th>WorldScore↑</th><th>cam-ctrl↑</th><th>3D-consist↑</th><th>photo↑</th></tr></thead><tbody> | |
| <tr><td class="lt mname">NVS-Solver</td><td><strong>70.44</strong></td><td>69.43</td><td><strong>91.63</strong></td><td>81.23</td></tr> | |
| <tr class="ours"><td class="lt mname">Ours · Wan2.1-VACE-14B <span class="tag ours">OURS</span></td><td>67.82</td><td>78.26</td><td>85.40</td><td>88.88</td></tr> | |
| <tr><td class="lt mname">FlexWorld</td><td>65.08</td><td><strong>79.48</strong></td><td>88.26</td><td><strong>89.60</strong></td></tr> | |
| <tr><td class="lt mname">Voyager · HunyuanVideo</td><td>63.29</td><td>74.99</td><td>80.46</td><td>83.46</td></tr> | |
| <tr><td class="lt mname">ViewExtrapolator</td><td>61.15</td><td>76.64</td><td>79.52</td><td>55.97</td></tr> | |
| <tr><td class="lt mname">TrajectoryCrafter</td><td>56.28</td><td>33.22</td><td>77.24</td><td>82.07</td></tr> | |
| <tr><td class="lt mname">GEN3C · Cosmos-7B</td><td>44.39</td><td>53.35</td><td>53.19</td><td>38.83</td></tr> | |
| <tr><td class="lt mname">FrameCrafter · Wan2.1-14B</td><td>41.14</td><td>50.95</td><td>37.22</td><td>0.00</td></tr> | |
| <tr><td class="lt mname">ViewCrafter</td><td>32.27</td><td>20.91</td><td>32.36</td><td>0.00</td></tr> | |
| </tbody></table></div> | |
| <div class="note">Ours is <b>2nd overall (67.8)</b> and <b>best-but-one on camera control (78.3)</b>; nvs_solver leads overall (70.4). WorldScore aggregates camera controllability, 3D consistency and photometric quality across world-generation prompts.</div> | |
| </section> |