Research & Innovation
Can a Neural Network Learn to Render 3D Gaussian Splats With No Rasterizer at All?
A new preprint under review at TMLR asks whether a transformer can render 3D Gaussian splats without a rasterizer, turning each Gaussian into a token an attention-based decoder reads directly into pixels. Trained on over 26,000 pruned Objaverse objects, it renders unseen objects at 38.08dB PSNR against rasterizing the same pruned input at 44.02dB — and the authors' own diagnostics point to exactly where that gap comes from.