Open-DiffusionGS is an open re-implementation and re-training of DiffusionGS, an ICCV 2025 paper titled "Baking Gaussian Splatting into Diffusion Denoiser for Fast and Scalable Single-stage Image-to-3D Generation and Reconstruction," authored by Yuanhao Cai together with researchers from Adobe and academic collaborators. The README is explicit that this codebase differs from the original version developed at Adobe: it is retrained from scratch on the open Objaverse/G-Objaverse dataset rather than Adobe's proprietary training data, and the released checkpoints are the authors' own reproduction. The core idea is a single-stage diffusion model that directly denoises 3D Gaussian primitives instead of relying on a separate multi-view diffusion model as an intermediate step, which the authors say lets it handle both object-centric single-view 3D generation (with mesh export) and scene-level reconstruction from a single image without a depth estimator, which the README says runs in about six seconds. Mesh exportation code was added to the repository in November 2025, so that feature ships today rather than being planned. The project provides four training scripts covering object and scene models at 256 and 512 resolution, an evaluation pipeline that reports PSNR/SSIM/LPIPS on the RealEstate10K benchmark (a released 256-resolution checkpoint scores 21.26 PSNR / 0.672 SSIM / 0.257 LPIPS), and pretrained weights plus a prompt-image dataset hosted on Hugging Face. In a qualitative comparison the authors ran themselves, they report their model completing inference in about 24 seconds versus roughly 180 seconds for Tencent's Hunyuan3D-v2.5, a claim from their own README rather than an independently verified benchmark. The repository is released under the MIT license, confirmed by reading the LICENSE.txt file directly (GitHub's own license detector also correctly flags it as MIT, unlike some repos where the detector misfires).