Bridging neural rendering and classical computer graphics

Recent advances in neural rendering have improved both the quality and efficiency of
novel view synthesis from multi-view images. In particular, 3D Gaussian Splatting
stands out by representing scenes with volumetric primitives and yielding high visual
fidelity in real time. Despite these successes, point-based primitives struggle with
sharp geometry, irregular surfaces, and integration with classical graphics pipelines.
This thesis addresses these limitations by investigating radiance field representations
based on alternative geometric primitives. The selection of a particular primitive sig-
nificantly affects reconstruction accuracy, rendering performance, and compatibility
with standard graphics pipelines. We introduce three complementary approaches
that progressively enhance visual quality and integration with graphics pipelines:
smooth convexes, triangles, and, finally, culminate in meshes.
First, we introduce 3D Convex Splatting, the first work to move beyond Gaussian
primitives by introducing smooth convex shapes. Compared to Gaussians, convex
shapes offer greater flexibility and more accurately represent flat surfaces and sharp
edges with fewer primitives. Coupled with an efficient CUDA rasterizer, it improves
fidelity while reducing memory usage.
Second, we revisit one of the most fundamental primitives in computer graphics:
the triangle. The main advantage of triangles is that they are fully compatible with
current rendering pipelines. Triangle Splatting develops a differentiable rendering
pipeline that enables direct optimization of triangles through end-to-end gradients.
It represents a promising first step toward mesh-aware neural rendering, bridging
decades of GPU-accelerated graphics with modern differentiable frameworks.
Finally, MeshSplatting extends this line of work by optimizing fully opaque, con-
nected triangles, jointly refining geometry and appearance through differentiable
rendering. This produces coherent surfaces that integrate seamlessly with real-time
rendering engines, enabling applications such as augmented and virtual reality without any postprocessing.
Together, these contributions show that alternative primitives provide a powerful
complement to Gaussian-based radiance fields. By progressing from convex shapes
to triangles and finally to meshes, this thesis bridges neural rendering and classical computer graphics, enabling high-quality reconstruction, geometric consistency,
efficient rendering, and full compatibility with real-time rendering engines.