Skip to main content
Back

LitePT Beats Point Transformer V3 at 3.6x Fewer Parameters and Twice the Speed

A new point cloud backbone uses convolutions early and attention late, cutting parameters 3.6x while matching Point Transformer V3 accuracy.

  • LitePT is a hybrid 3D point cloud backbone accepted as a CVPR 2026 Highlight.
  • Uses convolutions in early high-resolution stages and attention in deep low-resolution stages.
  • 3.6x fewer parameters, 2x faster, 2x less memory than Point Transformer V3.
  • Matches or beats PTv3 on ScanNet, nuScenes, Waymo, and Structured3D semantic segmentation.
  • Introduces PointROPE, a parameter-free 3D rotary positional encoding compatible with FlashAttention.
  • Code and weights available on GitHub and Hugging Face.

The CVPR 2026 Highlight paper LitePT paper reorganizes Point Transformer V3 around a simple finding: convolution and attention deliver their best returns at different depths of a point-cloud U-Net. LitePT assigns sparse convolutions to early, high-resolution stages and attention to deeper, lower-resolution stages. The resulting LitePT-S reportedly matches or exceeds PTv3 on several segmentation benchmarks with about 72% fewer parameters, twice the speed, and half the memory use.

Costs flip through the hierarchy

Point Transformer V3 processes sparse voxelized point clouds through a U-Net that progressively reduces spatial resolution before restoring it. Its stages combine sparse convolutions, which aggregate nearby geometry, with Transformer blocks that use attention and multilayer perceptrons to connect point features.

The authors attribute 67% of PTv3’s parameters to sparse convolution layers used for positional encoding. Attention and MLP components account for another 30%. Attention drives much of the latency because its pairwise computation scales quadratically within each attended token group, although hierarchical downsampling makes later stages cheaper by reducing the number of point tokens.

Those cost profiles move in opposite directions through the network. Sparse convolutions consume much of the parameter budget in deep stages, and attention is most expensive in early stages with many tokens. PCA projections from convolution-only, attention-only, and combined variants suggest that the hierarchy preserves the same functional progression: early features represent local geometry, and deeper features encode higher-level semantics.