TTB Employs Test‑Time Baking to Accelerate Decoder‑Only View Synthesis Rendering
training
| Source: HF Papers | Original article
Researchers introduce Test-Time Baking, a method that retains the speed benefits of test-time training while preserving the high-quality output of decoder‑only large view synthesis models with KV‑cache.
A paper posted to arXiv on 13 November 2026 introduces “Test‑time Baking” (TTB), a technique that dramatically speeds up decoder‑only large view synthesis models (LVSMs) while preserving their visual fidelity. LVSMs that rely on a key‑value cache have set the state‑of‑the‑art for novel‑view generation without building explicit 3‑D geometry, but their compute demands remain a bottleneck. Prior work has inserted test‑time training (TTT) layers to cut inference cost, only to sacrifice image quality.
TTB tackles this trade‑off by learning to bake global scene information extracted from cross‑view attention into a compact representation that can be reused at render time. The authors – Chin‑Yang Lin, Ryo Hachiuma and Jaesung Choe, working across NVIDIA and National Yang‑Ming Chiao Tung University – report that a decoder‑only view‑synthesis model that previously rendered at roughly one frame per second now runs at 23 FPS, with “near‑identical visual quality.” The speedup stems from eliminating the per‑frame attention passes that dominate the original pipeline.
The breakthrough matters because real‑time view synthesis underpins emerging AR/VR experiences, interactive gaming, and robotics perception, where latency and power budgets are tight. By retaining the high‑quality output of KV‑cached LVSMs while achieving near‑real‑time performance, TTB narrows the gap between research prototypes and production‑ready systems.
The community will now watch for broader evaluations on diverse datasets, open‑source releases of the baking modules, and integration into existing rendering stacks. Follow‑up work may explore extending the baking concept to other decoder‑only tasks such as video generation or physics‑aware simulation, potentially reshaping how large language‑style models are deployed in visual domains.
Sources
Back to AIPULSEN