InfiniSplat v2 cuts LPIPS by 25.8% in feed-forward 3DGS via frozen geometry branch
zhenjun_zhao · x · 2026-10-21
Researchers present InfiniSplat v2, analyzing why feed-forward Gaussian Splatting degrades under large viewpoint changes and fixing it with a frozen geometry branch decoupled from trainable Gaussian attribute prediction, scale-aligned GT camera supervision, and large-angle target view training. One trained model achieves state-of-the-art mean LPIPS/DISTS across six datasets in both pose-conditioned and pose-free settings, cutting LPIPS by 25.8% and DISTS by 32.9% vs the strongest baselines.
More from Research
- Embedding arithmetic demo: cartoon − animation − comic = diagram — mayfer · 2026-10-23
- Omnii Genome: Single Generalist Model Outperforms AlphaGenome Across All Benchmarks — JosephJacks_ · 2026-10-23
- Domingos Cites New Result Claiming RL Is 'Basically Useless' vs Supervised Learning — pmddomingos · 2026-10-23
- Word Arithmetic on Embeddings: "Bathroom Minus Toilet Equals Kitchen" Beats SAE Probing — mayfer · 2026-10-23
- Wolfram Opens Rule 30 Prizes to AIs: $30,000 for Three Open Problems — cephaloform · 2026-10-23
- Stanford and Databricks use AI agents to reproduce medical findings and test analytical robustness — ChengleiSi · 2026-10-23