LLLMWIKI
正文研究地图阅读入口元数据

来源笔记 · 更新于 2026-09-25

tttLRM快权重记忆驱动长序列高斯重建

阅读原论文

#near-cvpr-2025#3d-vision#gaussian-splatting#autoregressive#primary-source

一句话结论

CVPR 2026将已知相机图像序列编码为LaCT的fast weights，并以虚拟视角token查询解码显式高斯，统一16–64视的批量或因果在线重建；序列复杂度线性仍受快权重容量、预训练多卡成本与长视角推断时间约束（主文§3–4 pp.36584–36589、补充§A–B）。

输入与机制

有内外参的静态视频/多视物体图和射线编码，24个LaCT块用window attention结合TTT键值重建的内层快权重更新；虚拟视角只查询，以射线深度及像素参数解码高斯（主文pp.36584–36585 Fig.2/Eq.1–4）。AR模式按顺序小批吸收视图，用已更新快权重重解码整场；仅merge旧高斯会积累误差。训练目标是MSE+VGG感知目标图，场景用单目伪深度和opacity正则；离线TTT-LVSM预训练/相机标定是投入，test-time内层更新并非逐场对真实target RGB做反传（§3.3–3.5）。

主表、比较边界与消融

  • DL3DV-140 32视作者非AR 7.2s/PSNR25.07/LPIPS.215，Long-LRM独立32视模型1s/24.10/.254；逐场Scaffold-GS30K 16min/24.77/.205，作者PSNR略强而LPIPS略弱。AR 7.5s/24.31/.237；64视作者14.8s/25.95/.195，对Long-LRM3.7s/24.63/.243、逐场3DGS13min/26.55/.164、Scaffold16min/27.07/.173（主文p.36588 Table2）。T&T 64视作者20.31 PSNR/.274 LPIPS，对3DGS20.78/.205（同表）。
  • GSO 100个物体，512²/16视作者10个virtual view 0.8s/34.67 PSNR，对GS-LRM 16 virtual 2.5s/33.55；输出查询视角预算不同（p.36588 Table1）。32视predict&merge 21.50/.318、完整重解码23.63/.259，节省计算需画质代价；Muon及深度/opacity消融20.44→20.68→20.76 PSNR（p.36589 Tables4–5）。
  • 补充p.1固定记忆室外24.45 vs室内24.96、高频24.20 vs低频25.97；另经额外128视微调可达26.80 PSNR，属于额外训练。场景先144×256/80K、288×512/6K、540×960/5K+1K，AR再3K；训练64×A100 80GB，单卡表格推断计时不能省略预训练预算及伪深度/COLMAP（补充pp.1–2 §B.1）。

限制、代码、链接

DL3DV/COLMAP相机和预训练单目伪深度需分账；物体训练8输入/8目标从同15图独立抽取可能交叠，不能当严密视角分离（补充§B.2）。仓库含推断与模型代码，LICENSE.md为Adobe Research License，仅非商用研究教学；未运行GPU。详见raw/ingest/2026-09-25-tttlrm/analysis.md。三维高斯泼溅：表示、优化、几何与动态任务；iLRM：多视紧凑视角token反复吸收高分辨率图像；EcoSplat：按目标图元数条件化前馈高斯。

元数据
{
  "id": "2026-09-25-tttlrm",
  "type": "source",
  "title": "tttLRM：快权重记忆驱动长序列高斯重建",
  "status": "reviewed",
  "note_status": "reviewed",
  "created": "2026-09-25",
  "updated": "2026-09-25",
  "source_type": "paper",
  "venue": "CVPR 2026",
  "authors": "[Chen Wang, Hao Tan, Wang Yifan, Zhiqin Chen, Yuheng Liu, Kalyan Sunkavalli, Sai Bi, Lingjie Liu, Yiwei Hu]",
  "published_at": "2026-06-01",
  "ingested_at": "2026-09-25",
  "canonical_links": [
    "https://openaccess.thecvf.com/content/CVPR2026/html/Wang_tttLRM_Test-Time_Training_for_Long_Context_and_Autoregressive_3D_Reconstruction_CVPR_2026_paper.html",
    "https://openaccess.thecvf.com/content/CVPR2026/papers/Wang_tttLRM_Test-Time_Training_for_Long_Context_and_Autoregressive_3D_Reconstruction_CVPR_2026_paper.pdf",
    "https://openaccess.thecvf.com/content/CVPR2026/supplemental/Wang_tttLRM_Test-Time_Training_CVPR_2026_supplemental.pdf",
    "https://github.com/cwchenwang/tttLRM"
  ],
  "raw_entry": "raw/ingest/2026-09-25-tttlrm/",
  "topics": "[topics/3d-gaussian-splatting]",
  "entities": [],
  "claims": [],
  "questions": [],
  "tags": [
    "near-cvpr-2025",
    "3d-vision",
    "gaussian-splatting",
    "autoregressive",
    "primary-source"
  ]
}