Ultra3D. Yiwen Chen, Zhihao Li, Yikai Wang, Hu Zhang, Qin Li, Chi Zhang, Guosheng Lin. - High resolution 3D generation framework with high efficiency.- VecSet to generate coarse object layout.- **Part Attention** to refine, 6.7x speedup.- SOTA quality

VGGT-Long. Kai Deng, Zexin Ti, Jiawei Xu, Jian Yang, Jin Xie. - Extend VGGT to km-scale, unbounded outdoor env 3D reconstruction.- Cuhnk-and-Align, solving VGGT memory bottleneck.- SOTA on camera tracking and pointmap estimation.- efficient and scalable

Dens3R: A Foundation Model for 3D Geometry Prediction. - High quality performance 3D foundation model.- intrinsic invariance pointmap.- position-interpolated rotary positional encoding.- Enhancing a wide range of downstream tasks

SurfaceSplat: Connecting Surface Reconstruction and Gaussian Splatting. Zihui Gao, Jia-Wang Bian, Guosheng Lin, Hao Chen, Chunhua Shen. ZJU, Bytedance Seed, NTU. arXiv: Code(To be released):

SurfaceSplat: Connecting Surface Reconstruction and Gaussian Splatting. - SDF+3DGS to improve local fine details and global geometry coherence.- Improving sparse view recon. and NVS quality (5 - 10 inputs).- SOTA vs. Voxurf, 2DGS, GOF, etc. - ICCV 2025

PCR-GS: COLMAP-Free 3D Gaussian Splatting via Pose Co-Regularizations. Yu Wei, Jiahui Zhang, Xiaoqin Zhang, Ling Shao, Shijian Lu. - COLMAP free 3DGS SOTA.- DINO feature reprojection regularization.- Wavelet freq regularization.- ICCV 2025

GS-ID: Illumination Decomposition on Gaussian Splatting via Diffusion Prior and Parametric Light Source Optimization. Kang Du, Zhihao Liang, Zeyu Wang. HKUST, Tencent and Meituan.

GS-ID: - Realtime lightfield editing capability.- Large scale indoor/outdoor scene with realtime multi light source rendering.- Diffusion model prior for geometry/material estimation under unknown illumination.- Supports hybrid rendering of 3DGS and mesh

PI3: Scalable Permutation-Equivariant Visual Geometry Learning. - Breaking reliance on fixed reference view.- Permutation-equivariant.- SOTA for camera pose estimation, depth estimation, dense pointmap reconstruction .- Fast, robust, scalable.

SpatialTrackerV2 released its code and paper on arXiv(2507.12462). VGGT network + track refiner. First unified end-to-end framework for estimating camera motion, consistenct geometry and pixel-wise 3D tracking at once. Tracking up by 30%, dynamic reconstruction speed 50x.

PRoPE captures complete camera frustums, both intrinsics and extrinsics, as a relative positional encoding, enabling multi-view transformers with camera geometry to ground visual tokens in 3D space, improving various downstream tasks.

Second key component is StableNormal_turbo, which is trained with millions of high gloss, reflective and transparent material content, Code: Demo:

First key component is StableDelight, handling images with reflective material by segmenting non-Lambertian surfaces, and convert to Lambertian surfaces, Code: Demo: 在

Shout out to Kiri engine's 3DGS to Mesh v2.0, a collaboration with brilliant Hugo Ye, showcase at , also releasing two high quality open source project. Details in the thread

RT @ChrisHu996363: I have seen many inspiring papers to generate geometry with 3DGS like SuGAR, 2DGS, and so on. Now KIRI Engine brings the….

RT @Polycam3D: Big news!!. You can now make NeRFs with Polycam & @nerfstudioteam -- easy and free 🥳🎉. See the docs for more info: https://t….