sections/fairness.html

<section id="fairness">
<style>
#fairness { font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, Helvetica, Arial, sans-serif; color:#1a1a1a; line-height:1.5; }
#fairness h2 { font-size:1.6rem; margin:0 0 .4rem; border-bottom:2px solid #2b6cb0; padding-bottom:.3rem; }
#fairness h3 { font-size:1.15rem; margin:1.6rem 0 .2rem; color:#1a365d; }
#fairness .intro { background:#f0f6ff; border-left:4px solid #2b6cb0; padding:.7rem 1rem; border-radius:4px; margin:.6rem 0 1rem; font-size:.93rem; }
#fairness .blurb { font-size:.88rem; color:#444; margin:.2rem 0 .6rem; }
#fairness .grp { margin:.8rem 0 1.1rem; }
#fairness .grp-hdr { font-weight:600; font-size:.95rem; margin-bottom:.15rem; }
#fairness .ncommon { background:#2b6cb0; color:#fff; padding:.05rem .45rem; border-radius:10px; font-size:.85rem; }
#fairness .members { font-size:.8rem; color:#555; margin-bottom:.4rem; }
#fairness table { border-collapse:collapse; width:100%; font-size:.86rem; margin-bottom:.3rem; }
#fairness th, #fairness td { border:1px solid #d6dbe1; padding:.32rem .55rem; text-align:right; }
#fairness th:first-child, #fairness td:first-child { text-align:left; }
#fairness thead th { background:#1a365d; color:#fff; font-weight:600; }
#fairness tbody tr:nth-child(even) { background:#f7f9fc; }
#fairness tr.ours { background:#fff7e0 !important; font-weight:700; }
#fairness tr.ours td:first-child { color:#b7791f; }
#fairness .best { color:#1a7a3c; font-weight:700; }
#fairness .cov { font-size:.68rem; font-weight:400; color:#cbd5e0; }
#fairness .foot { font-size:.78rem; color:#777; font-style:italic; margin:.2rem 0 0; }
#fairness .arrow-note { font-size:.78rem; color:#666; margin-top:.4rem; }
</style>
<h2>Fair common-N comparison</h2>
<div class="intro"><b>Why this section?</b> Different methods were scored on different numbers of clips (different <i>N</i>) &mdash; some runs are partial, some methods skip clips they cannot handle &mdash; so the headline aggregates elsewhere are <i>not strictly comparable</i>. Here we fix that per benchmark: we take the <b>intersection of clip ids (cid) present in every scored method</b>, then <b>recompute each method&rsquo;s mean metrics over that identical clip set</b> and report <b>N_common</b>. Methods still scoring or with very few clips are marked <i>pending</i> and <b>excluded from the intersection</b> (so a half-finished run does not collapse N_common). Where a model does not produce reference-pixel metrics (it generates freely instead of following exact poses), we split into a <b>pixel-metric group</b> (exact-pose methods) and a <b>pose-free group</b> (geometry/perceptual metrics, all methods incl. Ours), each with its own N_common. All numbers below are recomputed from per-clip rows on disk &mdash; arrows mark higher-(&uarr;) or lower-(&darr;) better; <span class="best">green</span> = best, the <span style="color:#b7791f;font-weight:700">highlighted</span> row is Ours.</div>
<h3>Kubric-4D (synthetic dynamic NVS, perfect-GT anchor)</h3>
<p class="blurb">Exact-pose methods are scored with reference pixel metrics; our model and TrajCrafter generate freely (do not follow exact 81 poses) so they have only pose-free metrics. We therefore form two common-N groups.</p>
<div class="grp"><div class="grp-hdr">Pixel-metric group: <span class="ncommon">N_common = 90</span> (intersection of 6 methods)</div>
<div class="members">Methods (reference-pixel): FlexWorld, FrameCrafter, GEN3C, I2V-Base, NVS-Solver, ViewCrafter. Intersection gated on [psnr_full]; every aggregate recomputed on this identical clip set.</div>
<table><thead><tr><th>Method</th>
<th>PSNR&uarr;</th>
<th>SSIM&uarr;</th>
<th>LPIPS&darr;</th>
<th>PSNR-dyn&uarr;<br><span class="cov">cov &ge;0/90</span></th>
</tr></thead><tbody>
<tr><td>I2V-Base</td>
<td><span class="best">27.48</span></td>
<td><span class="best">0.7393</span></td>
<td>0.2782</td>
<td>&ndash;</td>
</tr>
<tr><td>FrameCrafter</td>
<td>26.78</td>
<td>0.6807</td>
<td>0.3183</td>
<td>10.98</td>
</tr>
<tr><td>NVS-Solver</td>
<td>25.21</td>
<td>0.7196</td>
<td><span class="best">0.2777</span></td>
<td><span class="best">14.00</span></td>
</tr>
<tr><td>ViewCrafter</td>
<td>25.01</td>
<td>0.7299</td>
<td>0.3256</td>
<td>11.52</td>
</tr>
<tr><td>GEN3C</td>
<td>19.48</td>
<td>0.7220</td>
<td>0.3419</td>
<td>11.75</td>
</tr>
<tr><td>FlexWorld</td>
<td>15.32</td>
<td>0.5904</td>
<td>0.4047</td>
<td>12.10</td>
</tr>
</tbody></table></div>
<div class="grp"><div class="grp-hdr">Pose-free group: <span class="ncommon">N_common = 60</span> (intersection of 7 methods)</div>
<div class="members">Methods (all incl. Ours): FlexWorld, FrameCrafter, GEN3C, NVS-Solver, TrajCrafter, VACE-14B (Ours), ViewCrafter. Intersection gated on [met3r, flow_warp_psnr]; every aggregate recomputed on this identical clip set.</div>
<table><thead><tr><th>Method</th>
<th>MEt3R&darr;</th>
<th>FlowWarp-PSNR&uarr;</th>
<th>COLMAP-ATE&darr;<br><span class="cov">cov &ge;37/60</span></th>
</tr></thead><tbody>
<tr class="ours"><td>VACE-14B (Ours)</td>
<td><span class="best">0.0143</span></td>
<td><span class="best">44.45</span></td>
<td>0.0774</td>
</tr>
<tr><td>GEN3C</td>
<td>0.0343</td>
<td>37.33</td>
<td>1.3646</td>
</tr>
<tr><td>FlexWorld</td>
<td>0.0428</td>
<td>40.80</td>
<td>0.2454</td>
</tr>
<tr><td>NVS-Solver</td>
<td>0.0633</td>
<td>29.76</td>
<td>2.1745</td>
</tr>
<tr><td>TrajCrafter</td>
<td>0.0900</td>
<td>30.52</td>
<td>0.8654</td>
</tr>
<tr><td>ViewCrafter</td>
<td>0.1076</td>
<td>25.42</td>
<td>0.3584</td>
</tr>
<tr><td>FrameCrafter</td>
<td>0.1125</td>
<td>24.56</td>
<td><span class="best">0.0003</span></td>
</tr>
</tbody></table></div>
<p class="foot">Pending (still scoring / too few clips, excluded from the intersection so they do not shrink N_common): Shape-of-Motion (N=15), Voyager (N=5).</p>
<h3>NVIDIA Dynamic Scenes (real dynamic, mono diagonal protocol)</h3>
<p class="blurb">Same split as Kubric-4D: pixel metrics over exact-pose methods; pose-free metrics (MEt3R, COLMAP-ATE) include our model and TrajCrafter.</p>
<div class="grp"><div class="grp-hdr">Pixel-metric group: <span class="ncommon">N_common = 18</span> (intersection of 9 methods)</div>
<div class="members">Methods (reference-pixel): FlexWorld, FrameCrafter, GEN3C, I2V-Base, NVS-Solver, Shape-of-Motion, ViewCrafter, ViewExtrapolator, Voyager. Intersection gated on [psnr_full]; every aggregate recomputed on this identical clip set.</div>
<table><thead><tr><th>Method</th>
<th>PSNR&uarr;</th>
<th>SSIM&uarr;</th>
<th>LPIPS&darr;</th>
<th>PSNR-dyn&uarr;<br><span class="cov">cov &ge;0/18</span></th>
</tr></thead><tbody>
<tr><td>FrameCrafter</td>
<td><span class="best">29.09</span></td>
<td>0.6780</td>
<td><span class="best">0.1622</span></td>
<td>13.62</td>
</tr>
<tr><td>I2V-Base</td>
<td>27.27</td>
<td>0.6407</td>
<td>0.3461</td>
<td>&ndash;</td>
</tr>
<tr><td>NVS-Solver</td>
<td>26.64</td>
<td><span class="best">0.6801</span></td>
<td>0.2892</td>
<td><span class="best">13.85</span></td>
</tr>
<tr><td>ViewCrafter</td>
<td>25.80</td>
<td>0.6219</td>
<td>0.3590</td>
<td>11.30</td>
</tr>
<tr><td>GEN3C</td>
<td>18.77</td>
<td>0.5559</td>
<td>0.3713</td>
<td>13.20</td>
</tr>
<tr><td>ViewExtrapolator</td>
<td>15.52</td>
<td>0.3462</td>
<td>0.5637</td>
<td>12.32</td>
</tr>
<tr><td>FlexWorld</td>
<td>14.89</td>
<td>0.3479</td>
<td>0.4851</td>
<td>12.00</td>
</tr>
<tr><td>Shape-of-Motion</td>
<td>14.89</td>
<td>0.4226</td>
<td>0.7855</td>
<td>10.71</td>
</tr>
<tr><td>Voyager</td>
<td>13.91</td>
<td>0.2932</td>
<td>0.6010</td>
<td>11.39</td>
</tr>
</tbody></table></div>
<div class="grp"><div class="grp-hdr">Pose-free group: <span class="ncommon">N_common = 18</span> (intersection of 10 methods)</div>
<div class="members">Methods (all incl. Ours): FlexWorld, FrameCrafter, GEN3C, I2V-Base, NVS-Solver, Shape-of-Motion, VACE-14B (Ours), ViewCrafter, ViewExtrapolator, Voyager. Intersection gated on [met3r]; every aggregate recomputed on this identical clip set.</div>
<table><thead><tr><th>Method</th>
<th>MEt3R&darr;</th>
<th>COLMAP-ATE&darr;<br><span class="cov">cov &ge;10/18</span></th>
</tr></thead><tbody>
<tr class="ours"><td>VACE-14B (Ours)</td>
<td><span class="best">0.0267</span></td>
<td>0.1330</td>
</tr>
<tr><td>FlexWorld</td>
<td>0.0673</td>
<td>0.5131</td>
</tr>
<tr><td>Voyager</td>
<td>0.0707</td>
<td>2.8937</td>
</tr>
<tr><td>I2V-Base</td>
<td>0.0977</td>
<td>2.9697</td>
</tr>
<tr><td>GEN3C</td>
<td>0.1050</td>
<td>1.5978</td>
</tr>
<tr><td>NVS-Solver</td>
<td>0.1063</td>
<td>0.6997</td>
</tr>
<tr><td>ViewCrafter</td>
<td>0.1123</td>
<td>0.3383</td>
</tr>
<tr><td>FrameCrafter</td>
<td>0.1230</td>
<td>0.7612</td>
</tr>
<tr><td>ViewExtrapolator</td>
<td>0.1239</td>
<td>0.6799</td>
</tr>
<tr><td>Shape-of-Motion</td>
<td>0.2834</td>
<td><span class="best">0.1046</span></td>
</tr>
</tbody></table></div>
<p class="foot">Pending (still scoring / too few clips, excluded from the intersection so they do not shrink N_common): TrajCrafter (N=14).</p>
<h3>SpatialVid (in-the-wild dynamic camera-trajectory set)</h3>
<p class="blurb">Our model follows exact poses on a subset of clips, so it appears in the pixel-metric group on those common clips; the pose-free group spans all methods including TrajCrafter.</p>
<div class="grp"><div class="grp-hdr">Pixel-metric group: <span class="ncommon">N_common = 10</span> (intersection of 8 methods)</div>
<div class="members">Methods (reference-pixel): FrameCrafter, GEN3C, I2V-Base, NVS-Solver, VACE-14B (Ours), ViewCrafter, ViewExtrapolator, Voyager. Intersection gated on [psnr_full]; every aggregate recomputed on this identical clip set.</div>
<table><thead><tr><th>Method</th>
<th>PSNR&uarr;</th>
<th>SSIM&uarr;</th>
<th>LPIPS&darr;</th>
<th>PSNR-dyn&uarr;<br><span class="cov">cov &ge;0/10</span></th>
</tr></thead><tbody>
<tr><td>FrameCrafter</td>
<td><span class="best">27.22</span></td>
<td><span class="best">0.7099</span></td>
<td><span class="best">0.2537</span></td>
<td><span class="best">16.83</span></td>
</tr>
<tr><td>I2V-Base</td>
<td>27.03</td>
<td>0.7096</td>
<td>0.3263</td>
<td>&ndash;</td>
</tr>
<tr><td>ViewCrafter</td>
<td>24.82</td>
<td>0.6481</td>
<td>0.3977</td>
<td>12.40</td>
</tr>
<tr><td>NVS-Solver</td>
<td>24.71</td>
<td>0.7059</td>
<td>0.3146</td>
<td>16.79</td>
</tr>
<tr><td>GEN3C</td>
<td>16.62</td>
<td>0.5925</td>
<td>0.4093</td>
<td>16.33</td>
</tr>
<tr class="ours"><td>VACE-14B (Ours)</td>
<td>15.53</td>
<td>0.4318</td>
<td>0.5736</td>
<td>15.85</td>
</tr>
<tr><td>ViewExtrapolator</td>
<td>14.70</td>
<td>0.4599</td>
<td>0.5535</td>
<td>14.74</td>
</tr>
<tr><td>Voyager</td>
<td>13.79</td>
<td>0.4977</td>
<td>0.5456</td>
<td>14.97</td>
</tr>
</tbody></table></div>
<div class="grp"><div class="grp-hdr">Pose-free group: <span class="ncommon">N_common = 14</span> (intersection of 9 methods)</div>
<div class="members">Methods (all incl. Ours): FrameCrafter, GEN3C, I2V-Base, NVS-Solver, TrajCrafter, VACE-14B (Ours), ViewCrafter, ViewExtrapolator, Voyager. Intersection gated on [met3r]; every aggregate recomputed on this identical clip set.</div>
<table><thead><tr><th>Method</th>
<th>MEt3R&darr;</th>
<th>COLMAP-ATE&darr;<br><span class="cov">cov &ge;13/14</span></th>
</tr></thead><tbody>
<tr><td>I2V-Base</td>
<td><span class="best">0.0608</span></td>
<td>0.5303</td>
</tr>
<tr><td>GEN3C</td>
<td>0.0613</td>
<td>0.7709</td>
</tr>
<tr><td>Voyager</td>
<td>0.0638</td>
<td>0.2891</td>
</tr>
<tr class="ours"><td>VACE-14B (Ours)</td>
<td>0.0711</td>
<td>0.1443</td>
</tr>
<tr><td>NVS-Solver</td>
<td>0.0787</td>
<td>0.3673</td>
</tr>
<tr><td>TrajCrafter</td>
<td>0.0898</td>
<td>0.4490</td>
</tr>
<tr><td>ViewCrafter</td>
<td>0.1001</td>
<td><span class="best">0.1211</span></td>
</tr>
<tr><td>ViewExtrapolator</td>
<td>0.1089</td>
<td>0.2257</td>
</tr>
<tr><td>FrameCrafter</td>
<td>0.1416</td>
<td>0.3547</td>
</tr>
</tbody></table></div>
<h3>DL3DV eval50 (50-clip static-scene extrapolation)</h3>
<p class="blurb">All methods follow exact poses except TrajCrafter; the pixel group covers exact-pose methods (incl. Ours), the pose-free group adds TrajCrafter.</p>
<div class="grp"><div class="grp-hdr">Pixel-metric group: <span class="ncommon">N_common = 34</span> (intersection of 9 methods)</div>
<div class="members">Methods (reference-pixel): FlexWorld, FrameCrafter, GEN3C, I2V-Base, NVS-Solver, VACE-14B (Ours), ViewCrafter, ViewExtrapolator, Voyager. Intersection gated on [psnr_full]; every aggregate recomputed on this identical clip set.</div>
<table><thead><tr><th>Method</th>
<th>PSNR&uarr;</th>
<th>SSIM&uarr;</th>
<th>LPIPS&darr;</th>
</tr></thead><tbody>
<tr><td>FrameCrafter</td>
<td><span class="best">26.48</span></td>
<td><span class="best">0.6586</span></td>
<td><span class="best">0.2365</span></td>
</tr>
<tr><td>I2V-Base</td>
<td>24.66</td>
<td>0.5968</td>
<td>0.3598</td>
</tr>
<tr><td>ViewCrafter</td>
<td>24.61</td>
<td>0.6160</td>
<td>0.3512</td>
</tr>
<tr><td>NVS-Solver</td>
<td>23.26</td>
<td>0.6302</td>
<td>0.3726</td>
</tr>
<tr><td>GEN3C</td>
<td>14.68</td>
<td>0.4444</td>
<td>0.5896</td>
</tr>
<tr class="ours"><td>VACE-14B (Ours)</td>
<td>14.48</td>
<td>0.4286</td>
<td>0.5201</td>
</tr>
<tr><td>FlexWorld</td>
<td>13.82</td>
<td>0.4052</td>
<td>0.5987</td>
</tr>
<tr><td>ViewExtrapolator</td>
<td>13.31</td>
<td>0.3611</td>
<td>0.6557</td>
</tr>
<tr><td>Voyager</td>
<td>11.54</td>
<td>0.3621</td>
<td>0.7170</td>
</tr>
</tbody></table></div>
<div class="grp"><div class="grp-hdr">Pose-free group: <span class="ncommon">N_common = 32</span> (intersection of 10 methods)</div>
<div class="members">Methods (all incl. Ours): FlexWorld, FrameCrafter, GEN3C, I2V-Base, NVS-Solver, TrajCrafter, VACE-14B (Ours), ViewCrafter, ViewExtrapolator, Voyager. Intersection gated on [met3r]; every aggregate recomputed on this identical clip set.</div>
<table><thead><tr><th>Method</th>
<th>MEt3R&darr;</th>
<th>COLMAP-ATE&darr;<br><span class="cov">cov &ge;21/32</span></th>
</tr></thead><tbody>
<tr><td>I2V-Base</td>
<td><span class="best">0.0529</span></td>
<td>1.9535</td>
</tr>
<tr><td>NVS-Solver</td>
<td>0.0772</td>
<td>1.3416</td>
</tr>
<tr><td>Voyager</td>
<td>0.1023</td>
<td>0.4459</td>
</tr>
<tr><td>FlexWorld</td>
<td>0.1057</td>
<td>0.4528</td>
</tr>
<tr class="ours"><td>VACE-14B (Ours)</td>
<td>0.1172</td>
<td>0.4089</td>
</tr>
<tr><td>FrameCrafter</td>
<td>0.1337</td>
<td>0.6157</td>
</tr>
<tr><td>GEN3C</td>
<td>0.1345</td>
<td>0.6805</td>
</tr>
<tr><td>ViewCrafter</td>
<td>0.1382</td>
<td><span class="best">0.1749</span></td>
</tr>
<tr><td>TrajCrafter</td>
<td>0.1883</td>
<td>0.7994</td>
</tr>
<tr><td>ViewExtrapolator</td>
<td>0.1987</td>
<td>0.5425</td>
</tr>
</tbody></table></div>
<h3>Long-Horizon DL3DV (extended-trajectory generation)</h3>
<p class="blurb">Reference pixels exist only on the GT-backed extrapolation clips (psnr_extrapGT), so the pixel group intersects over those clips; the pose-free group uses MEt3R / drift-ATE / collision over all clips. All scored methods follow exact poses here.</p>
<div class="grp"><div class="grp-hdr">Reference-pixel + pose-free group: <span class="ncommon">N_common = 9</span> (intersection of 11 methods)</div>
<div class="members">Methods (reference-pixel): FlexWorld, FrameCrafter, GEN3C, Lyra2, MatrixGame3, NVS-Solver, TrajCrafter, VACE-14B (Ours), ViewCrafter, ViewExtrapolator, Voyager. Intersection gated on [psnr_extrapGT]; every aggregate recomputed on this identical clip set.</div>
<table><thead><tr><th>Method</th>
<th>PSNR-extrapGT&uarr;</th>
<th>SSIM-extrapGT&uarr;</th>
<th>LPIPS-extrapGT&darr;</th>
</tr></thead><tbody>
<tr class="ours"><td>VACE-14B (Ours)</td>
<td><span class="best">13.89</span></td>
<td><span class="best">0.4248</span></td>
<td>0.6366</td>
</tr>
<tr><td>NVS-Solver</td>
<td>12.64</td>
<td>0.3949</td>
<td>0.7173</td>
</tr>
<tr><td>Lyra2</td>
<td>12.61</td>
<td>0.3708</td>
<td>0.7094</td>
</tr>
<tr><td>GEN3C</td>
<td>12.57</td>
<td>0.3795</td>
<td>0.6887</td>
</tr>
<tr><td>FlexWorld</td>
<td>12.55</td>
<td>0.4041</td>
<td>0.6826</td>
</tr>
<tr><td>FrameCrafter</td>
<td>12.33</td>
<td>0.3290</td>
<td><span class="best">0.6099</span></td>
</tr>
<tr><td>ViewCrafter</td>
<td>12.11</td>
<td>0.3511</td>
<td>0.6747</td>
</tr>
<tr><td>Voyager</td>
<td>11.55</td>
<td>0.3758</td>
<td>0.7398</td>
</tr>
<tr><td>TrajCrafter</td>
<td>11.40</td>
<td>0.3521</td>
<td>0.7530</td>
</tr>
<tr><td>MatrixGame3</td>
<td>11.27</td>
<td>0.3556</td>
<td>0.7565</td>
</tr>
<tr><td>ViewExtrapolator</td>
<td>10.63</td>
<td>0.2602</td>
<td>0.7728</td>
</tr>
</tbody></table></div>
<p class="foot">Pending (still scoring / too few clips, excluded from the intersection so they do not shrink N_common): VACE-14B-ft (Ours, LH-ft) (N=13).</p>
<p class="arrow-note">All aggregates in every table above are recomputed on the identical intersected clip set for that group &mdash; no method is averaged over clips another method in the group did not see.</p>
</section>