sections/glossary.html

<section id="glossary">
<style>
#glossary{font-family:-apple-system,BlinkMacSystemFont,"Segoe UI",Roboto,Helvetica,Arial,sans-serif;color:#1a1f29;margin:0 0 8px;}
#glossary h2{font-size:20px;margin:0 0 4px;font-weight:700;letter-spacing:-.01em;}
#glossary .gl-sub{font-size:12.5px;color:#5a6472;margin:0 0 10px;}
#glossary .gl-sub code{background:#f4f6f9;border:1px solid #e2e7ee;border-radius:4px;padding:0 4px;font-size:11.5px;}
#glossary .gl-legend{display:flex;flex-wrap:wrap;gap:7px 10px;margin:0 0 12px;font-size:11.5px;align-items:center;}
#glossary .gl-legend .lbl{color:#5a6472;font-weight:600;margin-right:2px;}
#glossary .tag{display:inline-block;padding:2px 8px;border-radius:11px;font-size:10.5px;font-weight:600;letter-spacing:.01em;white-space:nowrap;border:1px solid rgba(0,0,0,.08);}
#glossary .tag.warp{background:#dff0d8;color:#2c6a2f;}
#glossary .tag.ff{background:#dbe9fb;color:#1f5aa8;}
#glossary .tag.ar{background:#fde6c4;color:#9a5a00;}
#glossary .tag.gs3{background:#ece1f7;color:#6b3fa0;}
#glossary .tag.gs4{background:#f6dcef;color:#a3328e;}
#glossary .tag.mem{background:#d8f0ee;color:#1d6f68;}
#glossary .tag.tf{background:#fce4e4;color:#a83232;}
#glossary .gl-wrap{overflow-x:auto;border:1px solid #dfe4ea;border-radius:8px;}
#glossary table{border-collapse:collapse;width:100%;font-size:12px;background:#fff;min-width:880px;}
#glossary th,#glossary td{border:1px solid #e6eaef;padding:6px 9px;text-align:left;vertical-align:top;}
#glossary thead th{background:#eef1f5;font-weight:600;color:#222a35;position:sticky;top:0;white-space:nowrap;}
#glossary tr.grouphdr td{background:#dde3ec;font-weight:700;font-size:11.5px;color:#2a3340;letter-spacing:.01em;padding:6px 10px;}
#glossary td.name{font-weight:600;color:#202733;white-space:nowrap;}
#glossary td.desc{color:#33404f;min-width:240px;line-height:1.35;}
#glossary td.paper a{color:#1f5aa8;text-decoration:none;white-space:nowrap;}
#glossary td.paper a:hover{text-decoration:underline;}
#glossary td.paper .venue{display:block;font-size:10px;color:#7a8492;margin-top:1px;}
#glossary td.bb{color:#3a4452;font-size:11.5px;}
#glossary td.dash{color:#9aa3ad;}
#glossary tr.ours td{background:#fffdf3;}
#glossary tr.ours td.name{background:#fff4cf;box-shadow:inset 3px 0 0 #2c6a2f;}
#glossary tr:not(.grouphdr):hover td{background:#f6f9fc;}
#glossary tr.ours:not(.grouphdr):hover td.name{background:#fff0bf;}
#glossary .ours-badge{display:inline-block;margin-left:5px;font-size:9px;font-weight:700;color:#2c6a2f;background:#dff0d8;border-radius:3px;padding:1px 4px;vertical-align:middle;letter-spacing:.03em;}
</style>
<h2>Methods &mdash; glossary</h2>
<p class="gl-sub">All 20 methods in this benchmark, grouped by family. <b>Ours</b> rows (green accent) are warp+inpaint with Wan2.1-VACE-14B and its fine-tuned / I2V variants. arXiv ids are clickable; &ldquo;&mdash;&rdquo; = no verified paper (e.g. internal repro). Run ids in <code>monospace</code> match <code>runs/&lt;id&gt;_out/</code>.</p>
<div class="gl-legend">
<span class="lbl">Type tags:</span>
<span class="tag warp">warp+inpaint</span>
<span class="tag ff">feed-forward</span>
<span class="tag ar">autoregressive</span>
<span class="tag gs3">per-scene-3DGS</span>
<span class="tag gs4">per-scene-4DGS</span>
<span class="tag mem">world-model-memory</span>
<span class="tag tf">training-free</span>
</div>
<div class="gl-wrap"><table>
<thead>
<tr><th>Method</th><th>What it is</th><th>Paper / arXiv</th><th>Backbone</th><th>Type</th></tr>
</thead>
<tbody>

<tr class="grouphdr"><td colspan="5">Ours &mdash; warp + inpaint (Wan2.1-VACE)</td></tr>
<tr class="ours"><td class="name">Wan2.1-VACE-14B<span class="ours-badge">OURS</span><br><code>vace14b</code></td><td class="desc">Warp the depth point-cloud to target poses, then VACE inpaints the holes.</td><td class="paper"><a href="https://arxiv.org/abs/2503.07598">arXiv:2503.07598</a><span class="venue">VACE, 2025</span></td><td class="bb">Wan2.1-14B DiT (VACE)</td><td><span class="tag warp">warp+inpaint</span></td></tr>
<tr class="ours"><td class="name">VACE-14B ft (long-horizon)<span class="ours-badge">OURS</span><br><code>vace14b_ft_lh</code></td><td class="desc">VACE-14B with a long-horizon LoRA fine-tune for drift/loop trajectories.</td><td class="paper"><a href="https://arxiv.org/abs/2503.07598">arXiv:2503.07598</a><span class="venue">ours, FT of VACE</span></td><td class="bb">Wan2.1-VACE-14B + LoRA</td><td><span class="tag warp">warp+inpaint</span></td></tr>
<tr class="ours"><td class="name">VACE-14B ft (dynamic)<span class="ours-badge">OURS</span><br><code>vace14b_ft_dyn</code></td><td class="desc">VACE-14B fine-tuned on dynamic-scene warps for moving content.</td><td class="paper"><a href="https://arxiv.org/abs/2503.07598">arXiv:2503.07598</a><span class="venue">ours, FT of VACE</span></td><td class="bb">Wan2.1-VACE-14B + LoRA</td><td><span class="tag warp">warp+inpaint</span></td></tr>
<tr class="ours"><td class="name">VACE-14B ft (mixed)<span class="ours-badge">OURS</span><br><code>vace14b_ft_mixed</code></td><td class="desc">VACE-14B fine-tuned on a mixed static + dynamic + long-horizon corpus.</td><td class="paper"><a href="https://arxiv.org/abs/2503.07598">arXiv:2503.07598</a><span class="venue">ours, FT of VACE</span></td><td class="bb">Wan2.1-VACE-14B + LoRA</td><td><span class="tag warp">warp+inpaint</span></td></tr>
<tr class="ours"><td class="name">I2V base<span class="ours-badge">OURS</span><br><code>i2v_base</code></td><td class="desc">Wan2.1 image-to-video baseline conditioned on the first warped frame.</td><td class="paper"><a href="https://arxiv.org/abs/2503.20314">arXiv:2503.20314</a><span class="venue">Wan, 2025</span></td><td class="bb">Wan2.1-I2V-14B</td><td><span class="tag warp">warp+inpaint</span></td></tr>
<tr class="ours"><td class="name">I2V ft<span class="ours-badge">OURS</span><br><code>i2v_ft</code></td><td class="desc">Wan2.1-I2V fine-tuned on our warp+inpaint task as an ablation backbone.</td><td class="paper"><a href="https://arxiv.org/abs/2503.20314">arXiv:2503.20314</a><span class="venue">ours, FT of Wan-I2V</span></td><td class="bb">Wan2.1-I2V-14B + LoRA</td><td><span class="tag warp">warp+inpaint</span></td></tr>

<tr class="grouphdr"><td colspan="5">Diffusion baselines</td></tr>
<tr><td class="name">GEN3C<br><code>gen3c</code></td><td class="desc">3D-cache-conditioned video diffusion; renders a point-cloud cache to guide frames.</td><td class="paper"><a href="https://arxiv.org/abs/2503.03751">arXiv:2503.03751</a><span class="venue">CVPR 2025 Highlight</span></td><td class="bb">Cosmos / SVD video DiT</td><td><span class="tag warp">warp+inpaint</span></td></tr>
<tr><td class="name">Voyager<br><code>voyager</code></td><td class="desc">Long-range world-consistent RGBD video diffusion from one image + camera path.</td><td class="paper"><a href="https://arxiv.org/abs/2506.04225">arXiv:2506.04225</a><span class="venue">HunyuanWorld-Voyager, Tencent</span></td><td class="bb">HunyuanVideo DiT</td><td><span class="tag warp">warp+inpaint</span></td></tr>
<tr><td class="name">FlexWorld<br><code>flexworld</code></td><td class="desc">V2V diffusion + progressive 3DGS expansion for flexible 360&deg; view synthesis.</td><td class="paper"><a href="https://arxiv.org/abs/2503.13265">arXiv:2503.13265</a><span class="venue">FlexWorld, 2025</span></td><td class="bb">CogVideoX-5B-I2V &rarr; 3DGS</td><td><span class="tag gs3">per-scene-3DGS</span></td></tr>
<tr><td class="name">ViewCrafter<br><code>viewcrafter</code></td><td class="desc">Point-cloud-conditioned video diffusion for high-fidelity novel view synthesis.</td><td class="paper"><a href="https://arxiv.org/abs/2409.02048">arXiv:2409.02048</a><span class="venue">TPAMI 2025</span></td><td class="bb">DynamiCrafter video diff.</td><td><span class="tag warp">warp+inpaint</span></td></tr>
<tr><td class="name">FrameCrafter<br><code>framecrafter</code></td><td class="desc">Our LoRA reproduction / twin of VACE-14B trained on DL3DV.</td><td class="paper dash">&mdash;<span class="venue">ours, repro</span></td><td class="bb">Wan2.1-VACE-14B + LoRA</td><td><span class="tag warp">warp+inpaint</span></td></tr>
<tr><td class="name">NVS-Solver<br><code>nvs_solver</code></td><td class="desc">Zero-shot NVS: modulates SVD sampling with warped input views, no training.</td><td class="paper"><a href="https://arxiv.org/abs/2405.15364">arXiv:2405.15364</a><span class="venue">ICLR 2025</span></td><td class="bb">Stable Video Diffusion</td><td><span class="tag tf">training-free</span></td></tr>
<tr><td class="name">ViewExtrapolator<br><code>viewextrapolator</code></td><td class="desc">Refines artifact-prone rendered views via a redesigned SVD denoising; no fine-tune.</td><td class="paper"><a href="https://arxiv.org/abs/2411.14208">arXiv:2411.14208</a><span class="venue">2024</span></td><td class="bb">Stable Video Diffusion</td><td><span class="tag tf">training-free</span></td></tr>
<tr><td class="name">TrajectoryCrafter<br><code>trajcrafter</code></td><td class="desc">Redirects camera trajectory of a monocular video via dual-stream diffusion.</td><td class="paper"><a href="https://arxiv.org/abs/2503.05638">arXiv:2503.05638</a><span class="venue">ICCV 2025 Oral</span></td><td class="bb">CogVideoX-I2V</td><td><span class="tag warp">warp+inpaint</span></td></tr>

<tr class="grouphdr"><td colspan="5">Per-scene reconstruction</td></tr>
<tr><td class="name">Shape-of-Motion<br><code>shape_of_motion</code></td><td class="desc">Per-scene 4D-GS: fits SE(3) motion bases to one casual monocular video.</td><td class="paper"><a href="https://arxiv.org/abs/2407.13764">arXiv:2407.13764</a><span class="venue">2024</span></td><td class="bb">Dynamic 3D Gaussians</td><td><span class="tag gs4">per-scene-4DGS</span></td></tr>

<tr class="grouphdr"><td colspan="5">Long-horizon SOTA (world models)</td></tr>
<tr><td class="name">Lyra-2<br><code>lyra2</code></td><td class="desc">Feed-forward explorable 3D worlds; lifts diffusion video to 3DGS via self-distillation.</td><td class="paper"><a href="https://arxiv.org/abs/2604.13036">arXiv:2604.13036</a><span class="venue">Lyra 2.0, NVIDIA</span></td><td class="bb">Wan2.1-14B &rarr; 3DGS</td><td><span class="tag ff">feed-forward</span></td></tr>
<tr><td class="name">Matrix3D<br><code>matrix3d</code></td><td class="desc">All-in-one photogrammetry: pose, depth and NVS in one multi-modal diffusion model.</td><td class="paper"><a href="https://arxiv.org/abs/2502.07685">arXiv:2502.07685</a><span class="venue">2025</span></td><td class="bb">Multi-modal DiT</td><td><span class="tag ff">feed-forward</span></td></tr>
<tr><td class="name">SpatialMem<br><code>spmem</code></td><td class="desc">Video world model with geometry-grounded long-term spatial memory to fight forgetting.</td><td class="paper"><a href="https://arxiv.org/abs/2506.05284">arXiv:2506.05284</a><span class="venue">NeurIPS 2025</span></td><td class="bb">Video diffusion + mem.</td><td><span class="tag mem">world-model-memory</span></td></tr>
<tr><td class="name">VMem<br><code>vmem</code></td><td class="desc">Surfel-indexed view memory retrieves relevant past views for consistent generation.</td><td class="paper"><a href="https://arxiv.org/abs/2506.18903">arXiv:2506.18903</a><span class="venue">ICCV 2025 Highlight</span></td><td class="bb">Stable Video Diffusion + mem.</td><td><span class="tag mem">world-model-memory</span></td></tr>
<tr><td class="name">MatrixGame-3<br><code>matrixgame3</code></td><td class="desc">Real-time streaming interactive world model with long-horizon memory.</td><td class="paper"><a href="https://arxiv.org/abs/2604.08995">arXiv:2604.08995</a><span class="venue">Matrix-Game 3.0</span></td><td class="bb">WanX-derived DiT (DMD)</td><td><span class="tag ar">autoregressive</span></td></tr>

</tbody></table></div>
</section>