DensiTok densifies 3DGS tokens directly: 3 input views beat 16, faster
kwangmoo_yi · x · 2026-10-08
DensiTok, from Yonsei University with NAVER AI Lab and KIST, is a plug-in for pretrained feed-forward 3D Gaussian Splatting models. Key insight: sparse-view reconstruction lacks tokens, not pixels — instead of synthesizing extra RGB views and re-encoding them (costly, not 3D-consistent), it generates geometry tokens for unobserved views between the frozen encoder and reconstruction heads, compressing all token levels into a 4-channel latent completed in a single flow-matching step.
- Improves rendering at every sparse-view setting on AnySplat, WorldMirror and Depth Anything 3
- Standout result: AnySplat with DensiTok on 3 input views renders better and faster than AnySplat with 16 views
- Code coming soon
More from Multimodal
- AI film 'Gods Don't Give Gifts' hits theaters December 4, eyes Oscar animated feature bid — DavidmComfort · 2026-10-08
- Runway brings its video generation directly into ChatGPT: brief, generate, and give notes in one chat — runwayml · 2026-10-08
- 280+ generations to make Superman vs Saitama: MiniMax H3 workflow lessons — Automatic-Building84 · 2026-10-08
- Gossip Goblin Held Up as Proof AI Art Can Rise Above the Slop, With a Movie Trailer — PossibleVariety7927 · 2026-10-08
- 128,000 Micro-Dramas Launched in Q1 2026, Over 95% AI-Generated, Epoch AI Reports — Jsevillamol · 2026-10-08
- LiveAvatar Drops 30 New AI Avatars With 6 Emotions and Full-Body Streaming, Free Now — alifcoder · 2026-10-08