WaveView shapes diffusion noise phase, tops MetaView 0.84 dB
TL;DR
- WaveView gains 0.84 dB PSNR, 0.052 SSIM and 0.030 LPIPS over MetaView on DL3DV Hard's covisible region, and runs 2.1 times faster per view.
- The method injects a depth-warped source into the phase of a pretrained Qwen-Image-Edit generator's initial noise, releasing phase along the epipolar line where depth is uncertain.
- For dynamic NVS, fine-tuning AnyView on the same structured noise raises its zero-shot PSNR on AnyViewBench from 11.58 to 12.35 dB.
WaveView, a single-image novel view synthesis method from Karlsruhe Institute of Technology, beats the MetaView baseline by 0.84 dB PSNR, 0.052 SSIM and 0.030 LPIPS on the hard split of the DL3DV benchmark while running 2.1 times faster per view, according to the paper posted to Hugging Face. It adds no modules to the generator; it edits the initial Gaussian noise so the warped source controls the layout through the phase.
Jan-Hendrik Pauls, Yinzhe Shen, Frank Bieder and Christoph Stiller inject the depth-warped source into the phase of a pretrained Qwen-Image-Edit model using a dual-tree complex wavelet packet transform. Where the estimated depth is uncertain, phase is released along the epipolar line and kept across it; disoccluded regions stay pure noise. "WaveView keeps the architecture and adds negligible overhead," the abstract says, with depth estimation and noise construction adding only 0.46 seconds.
The same trick helps the dynamic case. Fine-tuning AnyView on the structured noise raises its zero-shot PSNR on AnyViewBench from 11.58 to 12.35 dB, with gains on all five datasets and a 1.57 dB jump on Argoverse. On a B200 GPU, WaveView takes 17.9 seconds per view, against 38.0 seconds for MetaView and 474.7 seconds for Voyager. It lands in a crowded week for diffusion-based 3D work, one of 72 computer-vision pieces we have tracked in the last 90 days.
The authors also fine-tune AnyView with plain Gaussian noise under the same recipe and reach only 11.35 dB PSNR, which the paper reads as evidence that the directional phase, not the extra training, is doing the work.
Originally reported by huggingface.co
Read the original article →Original headline: Karlsruhe WaveView Beats MetaView by 0.84 dB PSNR and Runs 2.1x Faster by Shaping the Phase of Noise