Test-Time Baking for Efficient Decoder-only View Synthesis
November 11, 2026
Test-Time Baking (TTB) uses a lightweight MLP to learn global scene information from cross-view attention. This method achieves the rendering efficiency of Test-Time Training (TTT) while maintaining higher fidelity for novel-view decoding.
HOW THIS AFFECTS YOU
●
builderYou can get faster, high-quality view synthesis without the heavy computation of full attention layers.
●
researcherYou can achieve high-fidelity rendering by repurposing fast-weight MLPs instead of replacing attention.