Congrong Xu

Excited to share our new work R³: 3D Reconstruction via Relative Regression. Only 372M params (~⅓ of recent 1B-class baselines), trained on 6×48G GPUs, but competitive on streaming reconstruction. Runs at 30+ FPS. Project: kevinxu02.github.io/r3-site/ Paper: arxiv.org/abs/2605.26519

Many thanks to my coauthors Huachen, @RoverXingyu @yuliangxiu @JunGao33210520 @AnpeiC and thank @siyuan_bian for insightful discussions!

Excited to share our new work R³: 3D Reconstruction via Relative Regression. Only 372M params (~⅓ of recent 1B-class baselines), trained on 6×48G GPUs, but competitive on streaming reconstruction. Runs at 30+ FPS. Project: kevinxu02.github.io/r3-site/ Paper: arxiv.org/abs/2605.26519

[4/4] The confidences also give us outlier rejection in a single pass — when a new frame's mean confidence against the active context drops below a baseline, we skip it. Handles occlusions, and scene cuts cleanly.

[3/4] But why does it work? The model doesn't have to assemble arbitrarily long sequences itself. Pairwise poses stay in-distribution, easier to retrieve past locations, and an easier prediction target.

Too many artifacts for GS reconstruction? Please checkout GenFusion: Closing the Loop between Reconstruction and Generation via Videos 🌐 Project page: genfusion.sibowu.com 💻 Code: github.com/Inception3D/Ge… #3D #DiffusionModels #ViewSynthesis #GenFusion #CVPR2025

Check out these amazing works also exploring Gaussians in the Wild! WildGaussians: wild-gaussians.github.io Wild-GS: arxiv.org/abs/2406.10373… GS-W: eastbeanzhang.github.io/GS-W/ SWAG: arxiv.org/pdf/2403.10427

Excited to share my first paper with @justkerrding and @akanazawa : Splatfacto in the Wild! Our method trains on a single RTX2080Ti and achieves real-time rendering! project page: kevinxu02.github.io/splatfactow/ paper: arxiv.org/abs/2407.12306

⚙️ Configure any model and generate commands with one click.

⏸️ Pause or stop your training at any time.

Excited to release 🚀Nerfstudio Gradio WebUI🚀: Simplify your 3D model training & visualization with an intuitive GUI! github.com/nerfstudio-pro… 📊 Train models directly from your browser 👁️ Visualize with Viser 📦 Process data seamlessly 🚀 Export models with ease And more...

Excited to announce 🚀gsplat v1.0🚀: a ⏩efficient⏩ CUDA backend for 3D Gaussian Splatting! docs.gsplat.studio A drop-in replacement of the official impl. with: - Up to 2x faster training; - Up to 4x less GPU memory; - Render millions of GSs in real-time; - And more;