Building a 3D scene from a few photos requires generating novel views. Top models store a global memory for high quality, but querying it for every pixel is slow. A faster approach skips the memory, but loses quality.
Test-Time Baking (TTB) takes a middle path. It bakes the key scene information from the global memory into a fast, lightweight neural network (a Multi-Layer Perceptron). This network answers novel-view queries instantly. A context pairing trick resolves a technical input mismatch between the baking and rendering stages, boosting quality.
TTB matches the quality of the best slow memory-based models while rendering at 23 frames per second. This is up to 23 times faster than previous methods, making high-quality interactive view synthesis practical.