Real Life Game Simulation Using 3D Gaussian Splatting and Stable Diffusion

Sarthak Sinha-21BPS1311 / Pushkar Raj-21BPS1150 / Priyanshu Kumar Pandey-21BLC1377

METHODOLOGY RESULTS
MOTIVATION / INTRODUCTION Experimental Setup - Experiments were conducted on
Constructing 3D models from 2D images is challenging due to 1) 3D Gaussian Splatting : Windows OS with 4GB of GPU storage on a GTX 1650 GPU set
the lack of depth information. Traditional photogrammetry up locally.
addresses this using multiple images and triangulation, but Sparse Point Cloud- Produce a sparse point cloud using
struggles with textureless or reflective surfaces. Advanced calibrated cameras and a set of photos of a static scene. 3D Gaussian Splatting Results - The 3D Gaussian
methods like Structure from Motion (SfM) and Multi-View Splatting was tested on multiple inputs, with parameters
3D Gaussians- From these points, characterized by position, adjusted to train the process on only 4GB of space. The
Stereo (MVS) improve accuracy by estimating depth and
camera positions. LIDAR provides precise geometric data, covariance matrix, and opacity, produce 3D Gaussians that training time was ~50 minutes
and when combined with photogrammetry, results in detailed,
allow for flexible optimization. Stable Diffusion Results - The Stable Diffusion results
textured 3D models.
Anisotropic Volumetric Splats- Anisotropic volumetric splats
were evaluated with standard metrics like peak-signal-to-
noise-ratio (PSNR) and structural similarity (SSIM).
OBJECTIVES
help produce a compact representation of the scene,
User-Friendly Solution-Create an accessible and
intuitive solution for game developers to bridge the gap especially for fine structures. CONCLUSION
between creative vision and technical implementation. This work introduces a pipeline aimed at assisting game
Spherical Harmonics- Spherical harmonics (SH) are used to developers in automating game settings and character
Effective Collaboration-Promote cooperation between game indicate directional appearance (color). creation from videos and images captured via a phone
developers and designers by offering strong communication camera. The proposed approach comprises three main
2) Incorporating Stable Diffusion : components: generating views from continuous frames,
tools and resources.
Image Smoothing-Effectively smooth images while integrating them into gaming engines such as Unity, and
Democratize Game Production-Remove barriers to employing Generative AI-based stable diffusion models to
preserving essential features such as edges, textures, and
entry for individuals with diverse backgrounds and skill create 3D meshes from single images.
sets, making ideation and wireframing more widely structural details.
available. REFERENCES
Robustness-Guarantee robustness against noise and
SCOPE OF THE PROJECT preserve image integrity even when parameters are changed ● Evaluating the Performance of Structure from Motion
Pipelines - MDPI
Encourage Data-Driven Design- Promote the use of
or the image is transformed.
● Photogrammetry: Step-by-Step Guide and Software
Synthetic Data-Contribute to the creation of realistic
●
Comparison - Formlabs
analytics and user feedback throughout the development cycle
Whitepaper on Structure from Motion (SfM)
to guide design decisions and improve gameplay experiences. and high-quality synthetic data.
Photogrammetry - ResearchGate
This objective ensures that games evolve based on real player
engagement and preferences.
Generative Gaussian Splatting-Synthesize new ● An Assessment of the Accuracy of Structure-from-Motion
(SfM) Photogrammetry for 3D Terrain Mapping -
perspectives from a single photograph of any item.
ResearchGate
Streamline Development WorkfloW- Optimize the ARCHITECTURE ● Using Differential Structure-from-Motion Photogrammetry
end-to-end game development process by integrating modular to Quantify Erosion Rates - ScienceDirect
tools and automation that reduce manual overhead. This ● Efficient NeRF Optimization - Not All Samples Remain
Equally Hard - arXiv
allows teams to focus more on creativity and problem-solving
● NeRF at CVPR23: Rendering Performance - Medium
rather than repetitive technical tasks, increasing productivity
and reducing time-to-market
● EfficientNeRF – Efficient Neural Radiance Fields - Jiaya
Jia
● 3D Gaussian Splatting for Real-Time Radiance Field
Support Cross-Platform Integration- Enable game Rendering - GitHub
creation that is adaptable across various platforms and
devices, ensuring seamless deployment to PC, mobile, VR, CONTACT DETAILS
and console environments. This flexibility empowers MAIL ID – [email protected],
developers to reach wider audiences without redundant [email protected] , [email protected]
redevelopment efforts. MOBILE NO- +917024318630, +917520241804 ,
+918889947761