UCSD's SPLATIFY Turns 3DGS Papers into Trainable Code, Cutting Weeks of Reimplementation to Minutes
UCSanDiego · hf · 2026-10-13
UC San Diego introduces SPLATIFY, a multi-agent framework that converts 3D Gaussian Splatting (3DGS) papers into trainable gsplat-based implementations where generic paper-to-code methods fail.
- A context-free grammar over a modular template constrains synthesis so code satisfies gsplat's architectural invariants by construction, with extension points for losses, densification, rendering and optimization.
- Faithful reproduction via fork-aware citation recovery, topologically ordered Graph-of-Thought synthesis, RAG-guided examples from 20+ verified implementations, and visual feedback (PSNR-guided regeneration, Gaussian-level checks, VLM patching).
- Compositional discovery improves PSNR by up to 2.4 dB over original results; cross-domain retrieval of physical priors yields novel methods like volumetric nebula rendering.
- Ships with SPLATIFY-Bench across 30 papers; matches expert reimplementations on code-less papers, cutting development from weeks to minutes.
More from coding & agent
- Indie dev ships hosted scraping MCP: pay nothing when sites block you, 1000 free credits — SidMishraDev · 2026-10-13
- Cursor's new pricing: Claude and other third-party models billed at API rates, plans include only Grok — iannuttall · 2026-10-13
- Dev: vibe coders with Claude can't yet beat 15-year principal engineers — but he expects yes within a year — facontidavide · 2026-10-13
- Open-Source MCP Security Gate Blocks Prompt Injections in Under 5ms — EstablishmentTough18 · 2026-10-13
- Open-source Claude skill writes working prompts on first try for 18+ AI tools — nikola_mr64990 · 2026-10-13
- When does an internal AI agent tool deserve to become a standalone app? — Moist_Deal_9995 · 2026-10-13