NAMVIS Swaps Diffusion for Autoregression, 3x Faster Views

A new arXiv preprint argues that sparse-view 3D synthesis does not need diffusion at all. The authors introduce NAMVIS, which they describe as 'a diffusion-free framework that reformulates multi-view image synthesis as geometry-conditioned next-scale autoregression.' The pitch is throughput: 'NAMVIS outperforms diffusion-based baselines in PSNR, SSIM, and LPIPS, while running over 3 times faster than the evaluated diffusion baselines under the same evaluation setting.'

The mechanism swaps iterative denoising for token prediction. NAMVIS 'predicts discrete visual tokens through a small number of coarse-to-fine scale steps, while sampling all tokens within each scale and across target views in parallel.' A component the paper calls Multi-scale Projective Pose Encoding 'injects source and target camera transformations into both target-view self-attention and source-to-target cross-attention at every resolution,' which is how the model keeps views geometrically aligned without the repeated denoising passes diffusion depends on.

The benchmarks are Objaverse, GSO, and OmniObject3D. The abstract names those three evaluation sets but does not identify which diffusion models the comparison runs against, and prints no absolute numbers for the PSNR, SSIM and LPIPS wins or for the per-view latency behind the 3x figure. The authors frame the finding as a direction rather than a settled benchmark: 'geometry-conditioned next-scale autoregression is a promising and efficient alternative to diffusion for sparse-view multi-view synthesis.'