Google Dec 2024 – Mar 2025 · ICCV 2025
Novel view synthesis at scale
Training large vision transformers to create new views of a scene, using generated examples to expand the available training data.
Co-first author
Novel-view reconstruction
Published reconstruction comparison
How it works
Training with generated views
CAT3D generates additional camera-conditioned observations. Tok-D separates source-image tokens from target-view tokens through layer-wise modulation, reducing the effect of artifacts in generated views on reconstruction.
Training and evaluation
The paper reports training on eight H100 GPUs for 100,000 iterations. Its real-data architecture comparison and synthetic-augmentation experiment are separate: the +1.64 dB DL3DV result is from the real-data comparison, while generated views produce smaller cross-dataset gains.
The architecture
Results
This table compares reconstruction quality after training on real data, evaluated at 256 × 256 resolution. PSNR measures image fidelity, with higher values indicating closer agreement with the reference image. Synthetic-data augmentation is a separate experiment in the paper.
| Dataset | LVSM PSNR | Tok-D PSNR |
|---|---|---|
| RealEstate10K | 28.89 dB | 30.02 dB |
| ACID | 29.19 dB | 29.47 dB |
| DL3DV | 19.91 dB | 21.55 dB |
ICCV 2025, Table 1. Real-data training, 256 × 256.
Higher PSNR is better. LVSM values are taken from its original paper.
This table compares the architectures after training on real images. The paper tests synthetic training data separately and finds smaller gains on unseen datasets. LVSM values come from its original paper.
Presentation
