LLLMWIKI
ArticleResearch mapReading portalMetadata

Source note · Updated 2026-09-25

PhysGMLarge Physical Gaussian Model for Feed-Forward 4D Synthesis

Single-image Gaussian and material prediction with synthetic preference supervision; physical ground truth remains unverified.

Read the paper

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

English translation pending — this page shows the Chinese original.

一句话结论

CVPR 2026用DINOv3/MVAdapter预训练与32×A800约三天训练，把单图映射成高斯与材料参数，随后MPM输出50帧；物性标注来自Qwen3-VL、参照视频由FramePack生成，表中视觉/偏好优势支持合成任务表现，真实材料参数精度仍需测量（主文pp.29857–29862，补充pp.1、3–8）。

输入、监督与机制

推断原始单图先由MVAdapter补三视，再用相机射线token+图像DINOv3特征进入24层Transformer，DPT输出每像素图元（零阶SH），三个全局token预测材料类别/E/ν条件分布；MPM位置/形变梯度更新图元，产生动态视角（主文Fig.2 Eq.1–8）。物理头DPO采三组参数，SAM-2/CoTracker-3比较合成视频与FramePack参照的视频轨迹，经边框/落地帧对齐标优劣（补充pp.7–8）。单图输出含预训练视图生成和MPM rollout的时间账。

主要证据与反证

  • 主文p.29861 Table1五材质平均CLIPsim .2748（w/DPO）/.2693（w/o）/.2291（DreamPhysics）；UPR 42.8%/30%/17.2%，补充p.2的四选一比较对象与主文Table1四行不同，100有效参与者×五场景的原始选择表需审计。分数及偏好均未直接验证真实力学。
  • p.29862 Table2 GSO 256² 本法25.47 PSNR/.071 LPIPS优于LGM21.44/.122；512²本法28.95 PSNR低于GS-LRM30.52，LPIPS .039优于.050。测试1,009个GSO对象，四视输入八视评测（补充p.1）；与单图+生成三视真实入口须分列。
  • p.29862 Table3场景生成<1min、DreamPhysics>0.5h；补充p.2又写本法<30秒。图1重建<1秒、模拟/渲染约30秒，MVAdapter时长、训练和标签生成另计。补充p.9承认刚性为主、整物体均匀材料、真实物性域差；沙类仅58样本（补充p.4 Table2）。

可用性与相关页面

CVF官方主文11页与官方ZIP补充18页、全文、逐表审计保存在raw/ingest/2026-09-25-physgm/；正式页未核到可用代码/权重许可，未GPU复现。3D Gaussian Splatting: Representation, Optimization, Geometry, and Dynamic Scenes将其与PhysGaussian：让高斯同时参与仿真与渲染按单图前馈/手设材料逐场的输入和成本分账。

Metadata
{
  "id": "2026-09-25-physgm",
  "type": "source",
  "title": "PhysGM：单图预测物性高斯并模拟4D动态",
  "title_en": "PhysGM: Large Physical Gaussian Model for Feed-Forward 4D Synthesis",
  "summary_en": "Single-image Gaussian and material prediction with synthetic preference supervision; physical ground truth remains unverified.",
  "status": "reviewed",
  "note_status": "reviewed",
  "created": "2026-09-25",
  "updated": "2026-09-25",
  "source_type": "paper",
  "venue": "CVPR 2026",
  "authors": "[Chunji Lv, Zequn Chen, Donglin Di, Weinan Zhang, Hao Li, Wei Chen, Yinjie Lei, Changsheng Li]",
  "published_at": "2026-06-01",
  "ingested_at": "2026-09-25",
  "canonical_links": [
    "https://openaccess.thecvf.com/content/CVPR2026/html/Lv_PhysGM_Large_Physical_Gaussian_Model_for_Feed-Forward_4D_Synthesis_CVPR_2026_paper.html",
    "https://openaccess.thecvf.com/content/CVPR2026/papers/Lv_PhysGM_Large_Physical_Gaussian_Model_for_Feed-Forward_4D_Synthesis_CVPR_2026_paper.pdf",
    "https://openaccess.thecvf.com/content/CVPR2026/supplemental/Lv_PhysGM_Large_Physical_CVPR_2026_supplemental.zip"
  ],
  "raw_entry": "raw/ingest/2026-09-25-physgm/",
  "topics": "[topics/3d-gaussian-splatting]",
  "entities": [],
  "claims": [],
  "questions": [],
  "tags": [
    "near-cvpr-2025",
    "3d-vision",
    "gaussian-splatting",
    "primary-source"
  ]
}