Apple Research Introduces Split-and-Scale Method to Expand 3D Latent Capacity at Inference

·2026.11.16 09:00

Key point

The method applies a single pretrained tokenizer to arbitrarily small sub-regions of a scene, allocating the same latent capacity to a smaller spatial extent to capture higher detail.

Details

  • Split-and-scale is a method that expands latent capacity at inference while keeping training cost tractable, addressing the high computational expense of scaling latents during training.
  • By representing 3D scenes as continuous functions of 3D coordinates, a single pretrained tokenizer is applied to small sub-regions, allowing the same latent capacity to cover a smaller spatial extent for higher detail.
  • Extensive ablations demonstrate that the inference-time scaling method achieves reconstruction quality competitive with much larger models that exceed single-GPU memory limits.
  • The approach supports high-quality generative modeling, with an autoregressive 3D generator trained using split-and-scale performing competitively with state-of-the-art baselines.
  • The research was published in November 2026 and presented at NeurIPS, authored by Dorian Chan, Muhammed Kocabas, Xiaoming Zhao, Oncel Tuzel, and Jen-Hao Rick Chang.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.