LLLMWIKI
ArticleResearch mapReading portalMetadata

Source note · Updated 2026-09-25

iLRM多视紧凑视角token反复吸收高分辨率图像

Read the paper

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

English translation pending — this page shows the Chinese original.

一句话结论

CVPR 2026把相机视角token分辨率与图像分辨率解耦，逐层按视角cross-attention吸收高分辨率证据，再跨视self-attention整合，得到较少高斯和可控多视推断成本；“迭代”指前馈层更新，不是逐场残差反传优化（主文§3 pp.37334–37336）。

输入和机制

输入已标定静态多视RGB、内外参；输出高斯及留出视角RGB。12层viewpoint token读取对应image token，token uplifting扩大低分辨率query，训练期quarter采样省跨注意力；最后线性回归高斯，无高阶SH；MSE+0.5感知损失监督留出相机（主文p.37336 Eq.4–5，补充p.1）。相机位姿需预处理与归一，独立几何GT不是主训练信号。

主表与反证

  • RE10K同两视和131072图元，iLRM28.65 PSNR/.110 LPIPS/0.025s对MVSplat26.39/.128/0.047s；4/8视结果30.37/31.57用了更多输入图，不能归因纯算法（主文p.37337 Table1）。
  • DL3DV 50帧、六视iLRM25.60 PSNR/.168 LPIPS/1.40GB，对DepthSplat24.19/.147/3.55GB，感知指标反向；12视512×960对照DepthSplat在4090上OOM，改H100运行（pp.37337–37338 Tables3–4）。
  • 32视540×960 H100，iLRM0.53s/24.30 PSNR，对逐场3DGS约8min/24.43；补充NRGBD 32视Chamfer .52劣于Long-LRM .43，F1 .69优于.59，几何指标不能单一排序（主文p.37338 Table5、补充p.3 Table4）。
  • 4视消融：每层重访图像29.24→单次cross28.58 PSNR，取消跨视self-attention→23.33；quarter cross训练迭代1.51→.94s、PSNR30.39→30.08，内存62.5→39.0GB在H100无checkpoint测（主文p.37339 Tables7–8、补充p.2 Table1）。256视推断20.92秒、相机平移扰动std .005时32视PSNR24.30→21.49，长期可扩展性与位姿依赖仍需测（补充p.4 Tables6–7）。

成本与相关页

2视训练8×H100 200K步、4视8×4090 100K、8视8×H100 50K；宽覆盖8×H200，完整卡时/相机准备未报。项目仓库有src但未声明许可证，复用需明确授权；未做GPU复现（补充p.1，GitHub API）。3D Gaussian Splatting: Representation, Optimization, Geometry, and Dynamic Scenes；MVSplat：匹配代价体驱动的前馈三维高斯。细节与口径见raw/ingest/2026-09-25-ilrm/analysis.md。

Metadata
{
  "id": "2026-09-25-ilrm",
  "type": "source",
  "title": "iLRM：多视紧凑视角token反复吸收高分辨率图像",
  "status": "reviewed",
  "note_status": "reviewed",
  "created": "2026-09-25",
  "updated": "2026-09-25",
  "source_type": "paper",
  "venue": "CVPR 2026",
  "authors": "[Gyeongjin Kang, Seungtae Nam, Seungkwon Yang, Xiangyu Sun, Sameh Khamis, Abdelrahman Mohamed, Eunbyung Park]",
  "published_at": "2026-06-01",
  "ingested_at": "2026-09-25",
  "canonical_links": [
    "https://openaccess.thecvf.com/content/CVPR2026/html/Kang_iLRM_An_Iterative_Large_3D_Reconstruction_Model_CVPR_2026_paper.html",
    "https://openaccess.thecvf.com/content/CVPR2026/papers/Kang_iLRM_An_Iterative_Large_3D_Reconstruction_Model_CVPR_2026_paper.pdf",
    "https://openaccess.thecvf.com/content/CVPR2026/supplemental/Kang_iLRM_An_Iterative_CVPR_2026_supplemental.pdf",
    "https://github.com/Gynjn/iLRM"
  ],
  "raw_entry": "raw/ingest/2026-09-25-ilrm/",
  "topics": "[topics/3d-gaussian-splatting]",
  "entities": [],
  "claims": [],
  "questions": [],
  "tags": [
    "near-cvpr-2025",
    "3d-vision",
    "gaussian-splatting",
    "neural-rendering",
    "primary-source"
  ]
}