NeurIPS Spotlight KoalaGS allocates Gaussians by image complexity for smooth LoD rendering
zhenjun_zhao · x · 2026-10-23
arXiv paper KoalaGS (Beche & de Charette, NeurIPS 2026 Spotlight) addresses continuous level-of-detail (LoD) in 3D Gaussian Splatting: deciding which Gaussians keep capacity as the rendering budget shrinks.
- A pilot study on 3DGS-MCMC reconstructions shows that image complexity, computed from input views before optimization, predicts where reconstruction error concentrates in image and 3D space; simple regions turn out to hold too many Gaussians and degrade as much as complex ones at low budgets.
- KoalaGS uses this complexity prior during training to bias Gaussian relocation toward poorly reconstructed areas while ensuring smooth degradation across LoDs; a primitive importance score then enables low-capacity inference while preserving scene structure.
- Across three datasets, it maintains competitive full-capacity quality while significantly improving quality across the budget spectrum, yielding smoother LoD transitions and faster progressive streaming.
More from Multimodal
- Paper turns an image generator into a generalizable dense correspondence matcher — kwangmoo_yi · 2026-10-23
- FreeMatching turns an image generator into a generalizable dense correspondence matcher — kwangmoo_yi · 2026-10-23
- How Claude edits videos: jacob_posel shares his skill-based workflow — jacob_posel · 2026-10-23
- PixAI open-sources Tagger 1.0, the 30,877-tag anime tagger behind Tsubaki.3's data curation — bdsqlsz · 2026-10-23
- FLUX Video Upscaler gets full upgrade, claims to beat Topaz Astra 2 at half the cost and 2x speed — bfl_ai · 2026-10-23
- Best local video generation models for 24GB VRAM? RTX 3090 user asks — b3nz1k · 2026-10-23