TTB: Test-time Baking for Efficient Rendering of Decoder-only View Synthesis Models
The authors report up to 23× faster rendering while matching state-of-the-art quality across several benchmarks.
Test-Time Baking trains a lightweight MLP to reproduce the cross-view attention mapping that carries global scene information. It retains attention’s guidance during this learning step, achieving the rendering efficiency of Test-Time Training with substantially higher quality, according to the paper. The method also separates camera-only input tokens from image-bearing target tokens so the MLP does not have to bridge the mismatch between context processing and novel-view rendering.
HF Daily Papers' note
Test-Time Baking trains a lightweight MLP to reproduce the cross-view attention mapping that carries global scene information. It retains attention’s guidance during this learning step, achieving the rendering efficiency of Test-Time Training with substantially higher quality, according to the paper. The method also separates camera-only input tokens from image-bearing target tokens so the MLP does not have to bridge the mismatch between context processing and novel-view rendering.
HF Daily Papers' note
score 5