FineSplat: Fine-Grained Patches Push Feed-Forward 3D Gaussian Splatting to New SOTA
zhenjun_zhao · x · 2026-10-23
FineSplat is a feed-forward 3D Gaussian Splatting framework built around fine-grained image patches, aiming to recover fine geometry and appearance that prior methods miss.
Key designs:
- Dual grouping strategy: view-wise grouping enables efficient interaction among neighboring views, while 3D spatial grouping lifts geometrically related tokens into 3D scene space to reconnect tokens that never interact under view-wise grouping — addressing the prohibitive cost of global multi-view attention at high token resolutions.
- Reconstruction error as direct feedback: an explicit rendering-based signal recovers fine-grained structures, overcoming one-pass feed-forward prediction limits.
- A compact Gaussian generation scheme preserves fine-grained features without increasing primitive count.
FineSplat achieves state-of-the-art results by a substantial margin under the same or smaller Gaussian budgets while remaining efficient and compact.
More from Research
- Berkeley's StyleStream 2.0 cuts streaming voice conversion latency to 520ms, 2x faster — Berkeley · 2026-10-23
- Math PhD student quits academia: AI is making math unrecognizable — pickover · 2026-10-23
- ICLR Submission Found Containing LLM's Own Meta-Comment Admitting Results Weren't Significant — _onionesque · 2026-10-23
- AKV Serving System Lets Agents Manage Their Own KV Cache, Cutting Cost 27.7% — _akhaliq · 2026-10-23
- AuroraTTS-100K: a sub-100K-parameter English TTS running 1120x real-time on GPU — Tall_Abrocoma_3533 · 2026-10-23
- MiTaS multi-resolution tactile imitation learning hits 80% success vs 35.5% vision-only baseline — Jan_R_Peters · 2026-10-23