Split-and-Scale Latent 3D Representations
Split-and-Scale Latent 3D Representations
The authors propose split-and-scale, an inference-time method that expands latent capacity by applying a single pretrained tokenizer to arbitrarily small sub-regions of a 3D scene, since representing scenes as continuous functions of 3D coordinates lets the same latent capacity cover a smaller spatial extent and capture far higher detail. They conduct ablations on trade-offs between compute, memory, latent capacity, and reconstruction quality, showing the method achieves quality competitive with much larger models that exceed single-GPU memory limits. They also train an autoregressive 3D generator with split-and-scale that performs competitively with state-of-the-art baselines.
Source: Apple Machine Learning Research · machinelearning.apple.com