NAMVIS: Next-Scale Autoregressive Multi-View Image Synthesis

NAMVIS. Next-Scale Autoregressive Multi-View Image Synthesis. By Ramil Khafizov and five others, from the Applied AI Institute, T-Tech, and one other institution. Published on the third of October. Its code repository is public and real: about forty source files.

Sparse-view novel view synthesis, producing new views of an object from only a few given ones, is a central problem in 3D content creation. Diffusion-based approaches are held back by their iterative denoising. Every generated view costs many denoising passes, which makes multi-view generation expensive at inference time.

NAMVIS drops diffusion entirely. It recasts multi-view image synthesis as next-scale autoregression, conditioned on geometry. Rather than denoising its way to each target view, it predicts discrete visual tokens through a small number of steps that go from coarse to fine. Within each scale, all tokens are sampled in parallel, and so are all the target views. The work happens in a handful of scale steps instead of a long chain of denoising.

This generation process has to be anchored to explicit camera geometry. For that the authors propose Multi-scale Projective Pose Encoding. It injects the source and target camera transformations into two places, the self-attention among target views and the cross-attention from source to target, and it does so at every resolution. On top of that, NAMVIS combines global conditioning with dense, geometry-aware cross-attention. The pairing is meant to let the model keep the appearance of the source views while keeping the target views consistent with one another.

Evaluation covers three datasets: Objaverse, GSO, and OmniObject3D. On all of them NAMVIS outperforms the diffusion-based baselines on PSNR, SSIM and LPIPS. It also runs more than three times faster than the evaluated diffusion baselines, under the same evaluation setting. So the comparison cuts both ways at once: better on all three quality metrics, and faster. The speed figure is the one the whole design was aimed at, a small number of parallel scale steps in place of iterative denoising, and the quality figures show that the geometry conditioning kept the views faithful while the steps were cut.

The authors draw a measured conclusion: geometry-conditioned next-scale autoregression is a promising and efficient alternative to diffusion for sparse-view multi-view synthesis. Note the word suggest. What is shown is a better result on three datasets in under a third of the time; what is proposed, from that, is a broader alternative to diffusion for this task.