Apple's Split-and-Scale Expands Latent 3D Capacity at Inference
Start here
Apple researchers propose split-and-scale, a method that expands the latent capacity of 3D representations at inference time while keeping training cost tractable. The approach applies a single pretrained tokenizer to arbitrarily small sub-regions of a scene, so the same latent capacity covers a smaller spatial extent and captures far higher detail. According to the researchers, the method matches the quality of much larger models that exceed single-GPU memory limits, and an autoregressive 3D generator trained on it performs competitively with state-of-the-art baselines.
Generated from reports · updated 1 h ago
Report timeline
Follow the reports to see the event from different sides.
- Apple Machine Learning ResearchApple Split-and-Scale Expands Latent 3D Representation Capacity at Inference
Apple researchers propose split-and-scale, a method that expands the latent capacity of 3D representations at inference time while keeping training cost tractable. By applying a single pretrained tokenizer to arbitrarily small sub-regions of a scene, it allocates the same latent capacity to a smaller spatial extent and captures far higher detail. The approach matches the quality of much larger models that exceed single-GPU memory limits, and an autoregressive 3D generator trained on it performs competitively with state-of-the-art baselines.
Heat for this event
Not enough continuous observations yet to draw a trend.