AI

NAMVIS: Next-Scale Autoregressive Multi-View Image Synthesis

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 08/10/2026, 11:18:24 UTC.

Briefings

NAMVIS is a diffusion-free framework for sparse-view multi-view synthesis that replaces iterative denoising with geometry-conditioned next-scale autoregression. It predicts discrete visual tokens via coarse-to-fine scale steps while sampling all tokens in parallel across views. The Multi-scale Projective Pose Encoding injects camera transformations into self-attention and cross-attention at every resolution. Evaluated on Objaverse, GSO, and OmniObject3D, NAMVIS outperforms diffusion baselines in PSNR, SSIM, and LPIPS while running over 3 times faster under identical settings.

Key points

• NAMVIS uses geometry-conditioned next-scale autoregression to replace iterative denoising in sparse-view synthesis.
• The Multi-scale Projective Pose Encoding injects camera transformations into both self-attention and cross-attention at every resolution.

Context

Diffusion-based approaches for multi-view image synthesis are limited by iterative denoising, making them expensive at inference time. NAMVIS offers a diffusion-free alternative that is faster and maintains high quality on multiple datasets.

Limitations

Results are limited to three datasets (Objaverse, GSO, OmniObject3D). Performance metrics include PSNR, SSIM, and LPIPS. No qualitative results, videos, or resources are included in the evidence.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 07/11/2026, 17:36:55 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeia5p5bnvul2ds7tgrd5zfwkruzabbpkshxzjwm5lxg7tyjncy5suy