Termsslug: split-and-scale

Split-and-Scale

拆分缩放

Method that splits scenes into small sub-regions at inference to expand latent capacity and detail.

Curated Intelligence

1 items · Page 1/1
Apple Machine Learning ResearchT1·Models
72 pts
Split-and-Scale Method Expands Latent Capacity at Inference

The paper introduces split-and-scale, which represents 3D scenes as continuous functions of coordinates so a single pretrained tokenizer can be applied at inference to arbitrarily small sub-regions, expanding latent capacity and detail without raising training cost. Extensive ablations validate the compute-memory-capacity-quality trade-offs, and the latent is further used to learn a generative model disentangling radiance fields and camera poses.