CompactSplat cuts 3D Gaussian counts 5x with texture-adaptive allocation for feedforward scenes
zhenjun_zhao · x · 2026-10-13
The arXiv paper CompactSplat tackles the redundancy of per-pixel uniform Gaussian allocation in feedforward 3D Gaussian Splatting with a content-aware framework:
- Texture-Aware Partitioner (TAP): a dual-path encoder (multi-view ViT + single-view ViT + U-Net) produces spatial masks by texture complexity
- Hierarchical Pyramid Sampler (HPS): dense Gaussians in complex regions; expanded, scale-modulated primitives tile flat areas
- Inference-time sparsity control: fidelity/memory trade-offs without retraining
Results: on RealEstate10K, PSNR up to 27.57 across S/B/L variants, with average Gaussians-per-pixel dropping from 3.0 to 0.554 and total Gaussians from 393.2k to 72.6k (5x fewer), while beating pixelSplat, MVSplat, DepthSplat and others on high-resolution rendering consistency (HRRC) and geometry.
More from Research
- Allen AI's TutorMoments Benchmark Finds LLM Tutors Over-Help Students — allen_ai · 2026-10-13
- Ben Goertzel revisits Omohundro's Basic AI Drives, finding six assumptions that decide which minds apply — bengoertzel · 2026-10-13
- Sim-to-real skepticism: pixel-space native training beats generated simulators, argues DeWitte — chris_j_paxton · 2026-10-13
- DeepMind publishes survey mapping theories of machine consciousness — Dr_Atoosa · 2026-10-13
- Noise: a speedrun for pretraining transformers without backprop — madhavsinghal_ · 2026-10-13
- Harvard releases freeinference_agentic_trace, a million-scale agent trace dataset — harvardMadsys · 2026-10-13