NVIDIA's TTB bakes scene info into an MLP, speeding view synthesis 23x
nvidia · hf · 2026-11-13
NVIDIA and NYCU researchers propose Test-Time Baking (TTB), a method for efficient novel-view rendering in decoder-only Large View Synthesis Models.
- Global scene information from cross-view attention is "baked" into a lightweight MLP, avoiding costly per-frame attention
- Maintains the efficiency of Test-Time Training approaches while improving novel-view fidelity
- Paper figures show up to 23x faster rendering than baselines like LVSM, reaching 960p at 23 fps
- Posted to arXiv on Nov 12, 2026
More from Multimodal
- Cinematic AI Short Film "Mechanical Metamorphosis" Shared on Reddit — ZidanStudio · 2026-11-13
- "All-in-One Tool for Minimax H3?" User Lists Demanding Video Workflow Wishlist — eeeeekzzz · 2026-11-13
- MiniMax H3-Image one-checkpoint 6-step edit demo runs on RTX 3090 in ~60s — Atagor · 2026-11-13
- From moodboard to 20-second teaser: a full coffee brand built in Adobe Firefly Boards and video editor — azed_ai · 2026-11-13
- Sgraffito Luminous Paradox: a reusable image-gen prompt template with striking results — LudovicCreator · 2026-11-13
- AI-generated cinematic "boyfriend" short goes viral with near-photoreal drama quality — Mollywhiteeee · 2026-11-13