麒砺创新技术（深圳）有限公司用户3321
Why 3D Gaussian Splatting Just Got Better with KIRI Engine Version 3.13 (Media Ver.)
输入“/”快速插入内容
Why 3D Gaussian Splatting Just Got Better with KIRI Engine Version 3.13 (Media Ver.)
Why 3D Gaussian Splatting Just Got Better with KIRI Engine Version 3.13 (Media Ver.)
用户4158用户41582025年1月23日修改
KIRI Engine Version 3.13: Optimizations All Around
KIRI Innovations has unveiled KIRI Engine Version 3.13, an update that introduces significant improvements to enhance functionality, efficiency, fidelity, and accessibility. With notable breakthroughs in 3D Gaussian Splatting technology, this update represents a pivotal step toward wider acceptance and practical implementation of 3DGS.
Higher Fidelity 3DGS - Improved Algorithm
The updated 3DGS algorithm offers significant improvements in visual quality. Users can expect enhancements in resolution, color accuracy, and rendering consistency across viewing angles. With the same input, the new algorithm produces 3DGS scenes with much higher fidelity, both when viewing the scene holistically and when examining details up close. These improvements open up more possibilities in practices such as field surveys, as well as environmental and design visualizations.
Results of KIRI's updated High Fidelity 3DGS (middle) and the 3DGS from two other scanner apps (left and right) using the same input.
This optimization draws inspiration from the Mini-Splatting2 project, by Guangchi Fang. Their research of aggressive Gaussian densification and visibility Gaussian culling laid the foundation for these advancements.
The optimization was achieved by increasing the input threshold, increasing the consideration of the source material's quality. It's natural to wonder whether the raised input threshold might impact generation speed or performance, but luckily, both concerns have been effectively addressed through additional optimizations.
Additionally, a new opening animation has been introduced, providing a visually engaging way to present projects. Embedding KIRI's 3DGS file viewer directly into web pages can boost user engagement by offering an attention-grabbing visual experience right from the start.
Faster Processing Times
The improved 3DGS algorithm now processes files in about half the time compared to previous versions. Given the same exact input, the previous version of 3DGS has an average of 5-7 minutes of Gaussian training duration, whereas the improved version completes the same task in 2-3 minutes. Cutting down the training time was achieved by reducing the number of points in the point-cloud data that is ply, focusing more on individual image resolution and visual presentation rather than the density of point-cloud representations. This improvement indicates enhancements and optimizations to the 3D Gaussian Splatting generation process overall, as it is able to reconstruct Gaussian splats with higher accuracy but also at a more efficient speed.
This reduced time opens up opportunities for near real-time environmental visual data to be transferred easily. This optimization also enables rapid generation, which is valuable for iterative workflows and can effectively allow for more trial and error.
File Size Optimization
Another key improvement is the reduction in 3DGS file sizes, which are now approximately 50% smaller. Because of the reduction in point-cloud density, the file size is naturally optimized to be smaller, yet it possesses higher fidelity due to the improved emphasis on resolution.