Generative Augmentation for 3D Gaussian Splatting

The Evolution of 3D Rendering

Unlike NeRFs (Neural Radiance Fields), which use Multi-Layer Perceptrons to infer scene details by casting rays, 3D Gaussian Splatting defines scenes as millions of independent 3D Gaussians. This provides significantly faster rendering capabilities.

My Research Focus

I explored the integration of Diffusion Models to augment sparse training views. By utilizing a generative model capable of predicting novel views from a limited set of anchors, I managed to synthesize additional training frames, drastically improving the output quality of the resulting 3D GS representation.

Technical Deep Dive

• Implemented in PyTorch.
• Rendered using custom CUDA kernels.
• Leveraged pre-trained Stable Diffusion architectures modified for multi-view consistency.