3DTurboQuant : Training-Free Near-Optimal Quantization for 3D Reconstruction Models

Jae Joong Lee · 2026

Read 3DTurboQuant : Training-Free Near-Optimal Quantization for 3D Reconstruction Models by Jae Joong Lee (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-05366.txt.

Description

Lee shows that shrinking 3D reconstruction models needs no per-scene training at all, overturning the assumption behind every current compressor. Because the bulky parameter vectors sit in a middling dimension range, one random rotation reshapes any input so its coordinates follow a predictable Beta law. That makes a precomputed, data-free quantizer nearly optimal, under three times the theoretical floor. His method also predicts in advance which parameters can shrink and by how much, and it compresses one splatting model more than threefold with tiny rendering loss.

Catalog record

CREATORJae Joong Lee
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE38228
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.05366

Browse the full library · About the Tower of Babel