SPLATIFY: Reproduce, Discover, Innovate! From Papers and Ideas to Trainable 3DGS Code
SPLATIFY turns 3D Gaussian Splatting papers into trainable gsplat implementations, including papers with no public code.
The paper describes a multi-agent system constrained by a gsplat grammar and modular template, with hooks for losses, densification, rendering, and optimization. It uses citation recovery, dependency-ordered synthesis, retrieved examples from verified implementations, and visual feedback to repair generated code. The authors say SPLATIFY matches expert implementations on no-code papers in SPLATIFY-Bench and cuts development time from weeks to minutes. They also report compositional improvements of up to 2.4 dB PSNR and generated methods for scientific scenes such as volumetric nebula rendering. HF Daily Papers' note
The paper describes a multi-agent system constrained by a gsplat grammar and modular template, with hooks for losses, densification, rendering, and optimization. It uses citation recovery, dependency-ordered synthesis, retrieved examples from verified implementations, and visual feedback to repair generated code. The authors say SPLATIFY matches expert implementations on no-code papers in SPLATIFY-Bench and cuts development time from weeks to minutes. They also report compositional improvements of up to 2.4 dB PSNR and generated methods for scientific scenes such as volumetric nebula rendering. HF Daily Papers' note
score 5