NVIDIA TTB Bakes View-Synthesis MLPs for 23x Faster Renders

A decoder-only view-synthesis model that rendered at 1 FPS now renders at 23. That is the headline number in TTB: Test-time Baking for Efficient Rendering of Decoder-only View Synthesis Models, posted to arXiv on 13 November 2026 by a team split between NVIDIA and National Yang Ming Chiao Tung University, with first author Chin-Yang Lin's work done during an NVIDIA internship.

The mechanism, in the authors' own words: TTB is "a strategy that learns to bake global scene information from cross-view attention into a lightweight MLP, enabling fast and high-fidelity novel-view decoding." The twist on prior Test-Time Training work is that the fast-weight MLP is trained to imitate cross-view attention instead of replacing it — "achieving the same rendering efficiency as TTT while delivering substantially higher quality." A second trick disentangles context tokens into camera-only rendering tokens and ground-truth tokens carrying image content, so the baked MLPs do not have to generalize across modalities at inference.

On DL3DV at 960p with 32 context views, the paper reports TTB at 31.52 dB PSNR versus DVSM's 31.58 dB, with rendering jumping from 1 FPS to 23 FPS; against LaCT the speedup is 2.3x at higher quality. Training ran on 64 A100 GPUs with a 340p→480p→720p→960p resolution curriculum, initialized from DVSM checkpoints. The abstract's own ceiling: "up to 23x rendering speedup with quality on par with the state-of-the-art across several benchmarks" — the "up to" and "on par" are theirs, not ours.

The paper lands in a busy week on our computer vision tracker, which also logged World Labs opening Atlas to the public the same day.