Neural Rendering News
The latest Neural Rendering coverage curated for game developers.
arXiv cs.GR details GS-PI, a method for turning Gaussian-splat assets into fully relightable PBR content without per-scene joint optimization. The approach separates appearance decomposition from geometry, then uses diffusion and …
arXiv cs.GR details CARA, a hash-encoding tweak that cuts collision waste in multiresolution image fitting. The method adapts per-level resolution and adds an invertible pixel-shuffle to spread spatial load more …
A new global illumination method packs light transport into a 13D Gaussian mixture over position, direction, normals, and material. The result is millisecond-scale, view-independent rendering with far less memory than …
Researchers have proposed CVQPG, a hologram representation that swaps the usual 2D Gaussian basis in splatting-style pipelines for quadratic phase functions. In tests, it improved reconstruction quality by 0.19 dB …
A new Gaussian splatting pipeline swaps bulky spherical harmonics for compact neural appearance codes, cutting per-primitive appearance storage from 192 bytes to 28. The result is a 1.3× faster optimization …
PointGT is a new point-based 3D representation that lets teams edit geometry and texture at the same time. The method targets a long-standing gap in neural reconstruction: photorealistic assets are …
A new self-supervised method called P-CORE aims to make point-based neural editing survive large non-rigid deformations without ground-truth views of the edited shape. It enforces surface consistency before and after …
A new 3D Gaussian Splatting method tackles gradient vanishing with truncated updates, aiming to make scene reconstruction more stable and less sensitive to initialization. The approach improves both static and …
MeshSplatBench benchmarks triangle-based neural rendering all the way from native optimization to Unity deployment. The study shows that rasterizable output alone is not enough for production use: topology, engine compatibility, …
Researchers have introduced an interactive way to deform 3D Gaussian Splatting scenes while keeping the rendered result consistent with the edited shape. The method targets a long-standing problem in 3DGS …
Neural texture compression can beat BCn on size, but shader-side divergence has been the tax. A new decoder design cuts that divergence by 25% to 52% and reaches up to …
A new intrinsic decomposition method tackles a long-standing misattribution problem in point-based 3D scene reconstruction. Intrinsic PAPR replaces translucent volume aggregation with proximity-aware point rendering so each surface point can …
AMD GPUOpen details a lightweight attention-based neural GI model that reconstructs off-screen indirect illumination from reflective shadow maps. The pitch is practical for teams chasing better bounce lighting without a …
MetaSapiens v2 pushes real-time neural rendering toward practical VR/AR use, claiming an order-of-magnitude speedup while preserving visual quality. The system combines foveation-aware pruning, peripheral-quality reduction, stereo warping, and a binocular …
A new skinning prior called cut-cell skinning aims to make neural weight generation faster and more robust for production meshes. It approximates volumetric geodesic distances with a graph-based method, avoiding …
A new hybrid BRDF aims to give real-time rendering a better middle ground between classic microfacet shading and heavier neural appearance models. By pairing a GGX-style base with a small …
A new rendering framework for time-varying implicit neural volumes reaches 30-40 FPS at 1024x1024 on an RTX 4090 while keeping timestep changes to about 1-2 ms. It replaces dense ray …
IRGS++ pushes Gaussian splatting beyond pretty rasterized views by adding transport-aware inverse rendering for visibility and indirect light. The method targets inter-reflections, glossy and metallic materials, and novel relighting, aiming …
A new holographic rendering pipeline turns multiplane images into wave-optics CGH fast enough for practical use, with reported speedups up to 250,000x over primitive-based methods. The approach keeps image quality …
ECoNGS brings compressive neural Gaussian splatting to volume visualization, aiming to cut the bloat of per-scene primitive representations. The system uses lightweight networks to predict editable splats from anchor points, …