Split-and-Scale Latent 3D Representations
Reported by Apple Machine Learning Research1 source ↓
Apple researchers describe split-and-scale, a method that expands 3D latent capacity at inference by applying a pretrained tokenizer to smaller scene regions. They report reconstruction quality competitive with larger models that exceed single-GPU memory limits.
Key points
- The method expands latent capacity at inference without scaling up training
- It applies a pretrained tokenizer to smaller spatial regions of a 3D scene
- The authors report reconstruction quality competitive with larger models that exceed single-GPU memory limits
- An autoregressive 3D generator using the method performed competitively with state-of-the-art baselines
Details
The method addresses limits in how much detail existing latent 3D representations can capture. The authors say increasing latent capacity during training can be costly and constrained by memory, while simply adding latents at inference does not use that capacity effectively.
Split-and-scale treats scenes as continuous functions of 3D coordinates, allowing a pretrained tokenizer to process smaller scene regions and use the same latent capacity over less space. The authors report ablation studies of compute, memory, latent capacity and reconstruction quality, and say the method supports an autoregressive 3D generator that performs competitively with state-of-the-art baselines.
Written from coverage as of
Images
- Apple Machine Learning Research
Stocks in this story
Change since open: 20 Nov to 21 Nov 2026
Last refresh