Splitting 3D scenes into sub-regions boosts detail without extra training memory

A new technique called split-and-scale reuses a single 3D tokenizer on subdivided regions, achieving reconstruction quality that previously required far larger models trained on massive GPUs.

Big Tech

Dorian Chan · Muhammed Kocabas · Xiaoming Zhao · Oncel Tuzel · Jen-Hao Rick Chang

Apple

Research Digest·13 November 2026·2 min read

Illustration: Splitting 3D scenes into sub-regions boosts detail without extra training memory

In one line

Split-and-scale applies a single trained 3D tokenizer to scene sub-regions, so latent capacity and detail scale up at inference without retraining.

Before you read this

Two ideas this paper builds on

test-time scalingAllocating extra computation during inference to improve output quality.114 papers →flow matchingA generative training method that learns an ordinary differential equation mapping noise to data by matching velocity fields along probability paths.169 papers →

Read it as

3D models are built from compact codes, but fine details need more detail than those codes can hold. Training bigger codes needs too much memory, and adding more codes at use time stops helping quickly.

The method splits a scene into small pieces and runs the same already trained compressor on each piece after rescaling its coordinates. The compressor spends the same amount of code on a smaller area, so each piece gets finer detail. Total code grows without retraining.

Reconstruction quality is close to much larger models that cannot fit on one graphics card, and the representation also powers a 3D generator. That matters because it gives high detail without expensive retraining.

How it works

before and after

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

test-time scaling cost-quality trade-off token efficiency flow matching