FastGS is the official implementation of the CVPR 2026 Highlight paper 'FastGS: Training 3D Gaussian Splatting in 100 Seconds', and it does exactly what the title claims. Released by the `fastgs` GitHub organization under the MIT license, the project has accumulated 1,100 stars and 102 forks by collapsing the historically minutes-to-hours training cost of a high-quality 3D Gaussian Splatting (3DGS) scene down to roughly 100 seconds on a single modern GPU, while keeping rendering quality on par with the canonical Inria reference implementation. For anyone working in real-time radiance fields, AR/VR capture, or photogrammetry pipelines, this is a step-change in iteration speed. ## The Problem FastGS Targets The original 3D Gaussian Splatting paper proved that millions of anisotropic Gaussians could deliver real-time novel-view rendering at quality competitive with neural radiance fields, but training time remained a bottleneck. A typical Mip-NeRF 360 scene takes 20 to 60 minutes on an A100 with the reference Inria code, and that is the loop a researcher or capture engineer pays every time they tweak data, masks, or hyperparameters. FastGS targets that loop directly. By restructuring the densification schedule, batching, and gradient accumulation, it reaches comparable PSNR and SSIM in roughly one fiftieth of the wall-clock time. ## How the 100-Second Number Holds Up The paper reports 100-second training on standard 3DGS benchmark scenes such as Mip-NeRF 360, Tanks and Temples, and Deep Blending, measured on a single high-end consumer or datacenter GPU. The acceleration comes from three orthogonal changes: a redesigned adaptive densification policy that avoids over-splitting in early iterations, a tile-aware backward pass that better saturates GPU memory bandwidth, and an aggressive use of mixed precision in the rasterizer's gradient path. The official benchmarks show PSNR within roughly 0.1 dB of the reference Inria pipeline at 50x the training speed, which is the core claim that earned the CVPR 2026 Highlight designation. ## A Drop-In Acceleration Framework The project is framed as a general acceleration framework, not a one-off implementation. FastGS exposes the same scene input format as the Inria reference (COLMAP outputs, sparse point clouds, image masks), and produces standard `.ply` Gaussian assets that load directly into existing viewers and editors such as SuperSplat and gsplat-based renderers. That compatibility is deliberate: teams with existing 3DGS toolchains can swap FastGS in as the training stage without rewriting their capture, post-processing, or distribution pipeline. ## CUDA Internals and Memory Profile FastGS is roughly half Python and half custom CUDA. The Python layer handles scene loading, optimizer scheduling, and logging, while the CUDA kernels implement the tile-based rasterizer and its backward pass. The memory profile during training is meaningfully lower than the reference implementation because intermediate gradient buffers are tiled rather than allocated per-Gaussian, which is what makes the framework viable on consumer 24GB GPUs for scenes that previously required A100-class hardware. The repository ships precompiled CUDA wheels for common Linux + PyTorch combinations to skip the often-painful build step. ## Practical Workflows For a photogrammetry studio, FastGS turns 3DGS training into a true interactive step. Capture a scene in the morning, run COLMAP, train the splat in under two minutes, inspect the result in a viewer, adjust masks, and retrain — all before lunch. For VR and AR product teams, the same acceleration makes per-product or per-room splats economically feasible at scale, where the historical training cost would have made batch processing the only viable mode. Academic groups have already used FastGS in follow-up papers as a strong baseline that future acceleration claims have to clear. ## Limitations FastGS is an acceleration of training, not a fundamental rewrite of the splatting representation. Scenes that are pathological for vanilla 3DGS — extreme transparency, view-dependent specularity at glancing angles, very thin geometry — remain pathological for FastGS at the same quality. The 100-second figure assumes a properly registered COLMAP sparse reconstruction; capture quality still dominates downstream splat quality, and FastGS does not solve that. The custom CUDA kernels also tie the framework to NVIDIA GPUs for now, with no ROCm path in the current release. ## Who Should Use FastGS FastGS is the right default for any team currently running the Inria reference 3D Gaussian Splatting pipeline who values iteration speed, for academic groups benchmarking new densification or pruning schemes (where FastGS is now the realistic baseline to beat), and for product teams building 3DGS-powered AR/VR experiences that need to fit splat training into a real CI/CD pipeline. Researchers exploring entirely new radiance-field representations may find vanilla Inria 3DGS easier to fork; FastGS shines once the underlying algorithm is fixed and the goal is to ship.