
Researchers at NVIDIA and National Yang Ming Chiao Tung University report a method that renders new camera views up to 23 times faster than their DVSM comparator in a tested configuration. Called Test-Time Baking (TTB), it shifts work into scene preparation so that subsequent images can be produced faster. The study appeared as an arXiv version 1 preprint on November 12, 2026; peer review is not verified.
Novel-view synthesis predicts what a scene would look like from another camera position using existing images. TTB requires those images’ camera poses and intrinsics. Decoder-only models such as DVSM retrieve scene information through cross-view attention, consulting a cache that grows with the input. TTB instead fits scene-specific multilayer perceptrons, or MLPs, to attention outputs during preparation, then uses their updated weights when rendering.
The method also pairs two representations of each context view: one containing only camera information, and another containing image and camera information. The camera-only representation learns to predict an image-informed attention target, matching the type of input available when rendering a new view. The standard configuration uses three inner optimization steps. This preserves attention’s contribution during preparation while avoiding repeated access to its growing cache during rendering.
The authors trained on DL3DV-10K and evaluated 140 held-out DL3DV scenes, alongside external datasets. With 64 context views at 960p, TTB rendered at 23 frames per second versus DVSM’s 1. But preparing the scene took 65 seconds versus 60 seconds: the reported 23-fold gain concerns rendering, not the complete pipeline. Standard TTB also retained a quality deficit: PSNR was 30.20 versus 31.11 dB, SSIM 0.897 versus 0.920, and LPIPS 0.131 versus 0.104. Higher PSNR and SSIM and lower LPIPS indicate closer agreement with reference images.
An optional variant groups nearby cameras and fits separate MLPs for each group. It reached 31.33 dB PSNR, 0.919 SSIM and 0.108 LPIPS in that setting, approaching DVSM across the metrics at the same reported rendering speed and preparation time. This improves capacity by storing more scene state. External results varied: on Hike with 64 views at 960p, the grouped variant exceeded DVSM on all three image metrics, but other configurations showed mixed outcomes.
Smaller controlled experiments supported the proposed components, though their setup differed from the main benchmarks. The evidence remains author-reported, without independent replication, confidence intervals or repeated-training uncertainty. Timing hardware and procedures are not specified in the supplied material; the use of A100 GPUs for training does not establish rendering hardware. Architecture and training differences also limit causal comparisons. Fixed MLP capacity can constrain quality, and attention keeps preparation expensive. The approach therefore appears most relevant when many views share one prepared scene; the benchmarks do not establish performance with unknown cameras or unrestricted unseen viewpoints.
Cover illustration generated with AI; conceptual, not research evidence.