People will be thrilled to learn that 2D Gaussian Splatting (2DGS) has officially been integrated into Nerfstudio's gsplat. This feature had been an open pull request since June, with several rounds of contributions. The result of this collaborative effort is native support for 2DGS in gsplat, with a low VRAM overhead of approximately 4GB.

Additionally, pre-compiled Wheels for both Linux and Windows have been released and can be accessed here.

Although Gaussian Splatting did not originally support fisheye lenses, that changes today with the merging of an implementation for Fisheye-GS. Now, datasets captured with fisheye lenses are fully supported in gsplat. Another major addition is Fused Differentiable SSIM from Taming 3DGS, which offers additional training speed improvements.

Another smaller feature that I really like is the inclusion of a new camera movement named Ellipse, which mimics the circular camera trajectory around a scene.

The full changelog for 1.4 can be found below and on the gsplat Github page:

Highlights

• Support 2DGS

• Camera model: perspective, orthogonal, fish-eye

• Start to support pre-built wheels (python 10)

What's Changed

• Update build workflow by @liruilong940607 in #369

• fix (apparant) typos in compression docs by @christian-helms in #370

• eval supports multi-GPU by @liruilong940607 in #371

• correct a typo to fix #373 by @jckhng in #383

• Add option to render camera trajectory as an ellipse. by @jefequien in #380

• Add an option to disable world space normalization in examples/simple_trainer.py by @JojiJoseph in #379

• Fix points projection when sampling patches for training by @nnop in #386

• fix reshape error when multi GPU bachsize > 1 not packed by @Ultized in #382

• Fix typo by @cr333 in #392

• Make download script work on Windows, too by @cr333 in #397

• Fused differentiable SSIM by @rahul-goel in #396

• 2DGS by @FantasticOven2 in #208

• citation.bib by @maturk in #400

• Implement Bilateral Guided Gaussian Splatting. by @jefequien in #291

• Bump fused-ssim commit by @liruilong940607 in #405

• Update rendering.py by @liruilong940607 in #407

• Update CITATION.bib by @cr333 in #410

• Make commands easier in README by @MrNeRF in #420

• Implement Fisheye-GS. by @jefequien in #398

• Publish precompiled wheels in Github release and add wheel index by @martinResearch in #365

• 1.4.0 by @liruilong940607 in #421

• fix workflow by @liruilong940607 in #422

• Workflow fix by @liruilong940607 in #423

These updates highlight the strength of the research pipeline built on top of Nerfstudio. Several exciting pull requests are still open within the codebase, and they will continue to be merged into the main branch.

Nerfstudio and gsplat continue to be permissibly licensed with an Apache 2.0 license for commercial purposes. If you have been thinking about contributing, do it!